Advanced Quantum Field Theory

Lent Term 2013

Hugh Osborn
Latex Lecture notes, originally typeset by
Steen Gielen in 2007
revised by
Carl Turner in 2013
latest update: June 30, 2013
Books
There are many books on quantum eld theory, most are rather long. All those listed are worth
looking at.
M.E. Peskin and D.V. Schroeder, An Introduction to Quantum Field Theory
842p., Addison-Wesley Publishing Co. (1996).
A good introduction with an extensive discussion of gauge theories including QCD and
various applications.
M. Srednicki, Quantum Field Theory
641p., Cambridge University Press (2007).
A comprehensive modern book organised by considering spin-0, spin-
1
2
and spin-1 elds
in turn.
S. Weinberg, The Quantum Theory of Fields
vol. I Foundations, 609p., vol. II Modern Applications, 489p., Cambridge University Press
(1995,1996).
Written by a Nobel Laureate, contains lots of details which are not covered elsewhere,
perhaps a little idiosyncratic and less introductory than the above. There is a third
volume on supersymmetry.
J. Zinn-Justin, Quantum Field Theory and Critical Phenomena, 4th ed.
1054p., Oxford University Press (2002).
Devotes a large proportion to applications to critical phenomena in statistical physics but
covers gauge theories at some length as well; not really an introductory book.
C. Itzykson and J-B. Zuber, Quantum Field Theory
705p., McGraw-Hill International Book Co. (1980).
At one time the standard book, contains lots of detailed calculations but the treatment
of non abelian gauge theories is a bit cursory and somewhat dated.
T. Banks, Modern Quantum Field Theory, A Concise Introduction
271p., Cambridge University Press (2008).
As it says quite concise, contains an interesting selection of subjects and useful for sup-
plementary reading.
L. Alvarez-Gaume and M.A. Vazquez-Mozo, An Invitation to Quantum Field Theory
294p., Springer (2012).
Rather introductory.
M. Shifman, Advanced Topics in Quantum Field Theory
622p., Cambridge University Press (2012).
As the title indicates not an introduction but contains material on non perturbative
approaches.
There are also many much more mathematical approaches to quantum eld theory, many very
sophisticated.
J. Dimock, Quantum Mechanics and Quantum Field Theory: A Mathematical Primer
283 p., Cambridge University Press (2011).
Some more mathematical stu than usual.
P. Deligne et al., Quantum Fields and Strings: A Course for Mathematicians, vol. I,II
1499p., ed. P. Deligne, P. Etingof, D.S. Freed, L.C. Jerey, D. Kazhdan, J.W. Morgan,
D.R. Morrison, E. Witten, American Mathematical Society (1999).
Contains various articles by world renowned physicists and mathematicians, at a level
from the almost trivial to the sublime to the incomprehensible. A very alternative view.
K. Fredenhagen and K. Rejzner, Perturbative algebraic quantum eld theory
arXiv:1208.1428
A review of the mathematical approach to QFT.
i
Online material
W Siegel, Fields
885p., hep-th/9912205 or http://de.arxiv.org/pdf/hep-th/9912205.
Contains additional material on GR, Strings, Supersymmetry, Supergravity, an unusual
point of view.
S. Coleman, Notes from Sidney Colemans Physics 253a
337p., arXiv:1110.5013.
Notes from Sidney Colemans lectures at Harvard in 1986. He was renowned for his
lecturing, the notes are very introductory.
L. Alvarez-Gaume and M.A. Vazquez-Mozo, Introductory Lectures on Quantum Field
Theory
hep-th/0510040.
Rather elementary.
For more specialised material see
M. Polyak, Feynman Diagrams for Pedestrians and Mathematicians
arXiv:math/0406251.
I found the rst part instructive.
M. Flory, R.C. Helling and C. Sluka, How I Learned to Stop Worrying and Love QFT
23p., arXiv:1201.2714.
Some nice remarks on QFT but there are quite a few typos.
D. Harlow, A Simple Bound on the Error of Perturbation Theory in Quantum Mechanics
arXiv:0905.2466.
Has a simple discussion of convergence of perturbation theory.
E. Witten, Notes On Supermanifolds and Integration
arXiv:1209.2199.
A fairly accessible discussion of integrals over Grassman variables.
Specialist topics
J.C. Collins, Renormalization
380p., Cambridge University Press (1984).
The introductory chapters and discussion of dimensional regularization are good, later
chapters are rather technical and perhaps better covered elsewhere.
H. Kleinert, Path Integrals in Quantum Mechanics, Statistics, Polymer Physics, and Finan-
cial Markets, 3rd ed.
1468p., World Scientic (2003).
Not a eld theory book but the bible on path integrals.
For an historical perspective it is instructive to read
S.S. Schweber, QED and the Men Who Made It: Dyson, Feynman, Schwinger and Tomon-
aga
732p., Princeton University Press (1994).
This is a history of how Feynman, Schwinger and Tomonaga learned how to calculate in
quantum eld theory, and Dyson showed how Feynman rules could be derived.
L. ORaifeartaigh, The Dawning of Gauge Theory
249p., Princeton University Press (1997).
Describes the rather tortuous route to understanding gauge invariance.
ii
D.J. Gross, Oscar Klein and Gauge Theory
hep-th/9411233.
S. Weinberg, What is Quantum Field Theory, and What Did We Think It Is?
hep-th/9702027.
iii
Contents
1 Path Integrals 1
1.1 Standard Approach to Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . 1
1.2 Path Integral in One-Particle Quantum Mechanics . . . . . . . . . . . . . . . . . 2
1.2.1 Example: The Harmonic Oscillator . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 A Few Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Gaussian Integrals and Extensions Over Multi-Dimensional Coordinates . . . . . 10
1.4 Non-Gaussian Integrals and Perturbation Expansions . . . . . . . . . . . . . . . 14
2 Functional Methods in Quantum Field Theory 18
2.1 Free Scalar Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Interacting Scalar Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Feynman Rules for Interacting Scalar Field Theory . . . . . . . . . . . . . 22
2.2.2 Connected and Disconnected Graphs . . . . . . . . . . . . . . . . . . . . . 25
2.2.3 One Particle Irreducible Graphs . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.4 Tree Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Fermionic Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.1 Gaussian Integrals for Grassmann Variables . . . . . . . . . . . . . . . . 33
2.3.2 The Fermionic Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.3 Path Integral Formalism for Fermionic Fields . . . . . . . . . . . . . . . 35
2.3.4 Canonical Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.5 Propagators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Operator Aspects of Quantum Field Theory 39
3.1 Symmetries and Noethers theorem . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Ward Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 Energy Momentum Tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Relation of n-Point Functions and Scattering Amplitudes . . . . . . . . . . . . . 42
4 Calculation of Feynman Amplitudes and Divergences 46
4.1 Two Key Ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.1.1 Wick Rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.1.2 Appearance of Divergences in Feynman Integrals . . . . . . . . . . . . . 47
4.2 Regularisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.2.1 Dimensional Regularisation for One-Loop Graphs . . . . . . . . . . . . . . 51
4.2.2 Dimensional Regularisation for Two-Loop Graphs . . . . . . . . . . . . . 57
4.3 Renormalisation and the Renormalisation Group . . . . . . . . . . . . . . . . . . 61
4.3.1 Bare Lagrangians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.2 The Callan-Symanzik Equation . . . . . . . . . . . . . . . . . . . . . . 63
4.3.3 Evolution of Coupling Constants . . . . . . . . . . . . . . . . . . . . . . . 66
4.4 Eective Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5 Gauge Theories 72
5.1 Brief Summary of Lie Algebras and Gauge Fields . . . . . . . . . . . . . . . . . . 72
iv
5.2 Gauge Fields in Quantum Field Theory . . . . . . . . . . . . . . . . . . . . . . . 74
5.3 Example: Ordinary Finite Integral . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.4 Introduction of Ghost Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.5 Feynman Rules for Gauge Theories . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.6 Canonical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.7 BRS Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.8 Renormalisation of Gauge Theories . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.8.1 Calculation of (g) at One Loop . . . . . . . . . . . . . . . . . . . . . . . 91
6 Fermion Currents and Anomalies 98
6.1 Example: Symmetries of the Dirac Lagrangian . . . . . . . . . . . . . . . . . . 98
6.2 Triangle Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
v
Introduction
The course Advanced Quantum Field Theory will build on the course Quantum Field The-
ory taught in Michaelmas Term. It will extend the material covered in this course to interacting
theories (including loops) and more realistic theories, which can at least potentially predict ex-
perimental results. It will also introduce how to deal with gauge theories.
The basic message that this course tries to convey is:
Quantum eld theory is the basic language of particle physics, and also large parts of statistical
physics.
Quantum eld theory is a subject with many technical complications; we will try to deal
with these step by step. It is, however, not a branch of mathematics yet. The lectures will
not be rigorous from a pure mathematical point of view.
In quantum eld theory, the number of particles involved is potentially innite, whereas
ordinary quantum mechanics deals with states describing one particle or a xed number of
particles. Quantum eld theory is quantum mechanics with an innite number of degrees of
freedom. We are dealing with elds (x, t) dened on space-time. Quantisation of free elds
denes a space of states, or Fock
1
space:
a vacuum state [0), which has zero energy,
single-particle states [ p), which have energy E( p) =
_
p
2
+m
2
,
multi-particle states [ p
1
, p
2
, . . . , p
n
), which have energy E( p
1
) +. . . +E( p
n
).
(Note that we use units in which c = = 1.)
We introduce creation and annihilation operators a

( p), a( p), such that [ p) = a

( p)[0) etc.
Fields after quantisation change the number of particles. Whereas the number of particles is
conserved in the free theory, interactions can change the number of particles.
Quantisation takes you from elds to particles. (For example, quantising the electromagnetic
eld leads to photons with energy E = [ p[.)
1 Path Integrals
In this rst chapter, a dierent approach to quantum mechanics will be presented, the path
integral approach. We start with ordinary quantum mechanics, and the formalism will generalise
to quantum eld theory.
1.1 Standard Approach to Quantum Mechanics
We start from a classical Lagrangian
2
describing a system with n degrees of freedom q
i
, i =
1, . . . , n
L(q
i
, q
i
) . (1.1)
In the Hamiltonian
3
formalism, we replace q
i
by the conjugate momenta p
i
p
i
=
L
q
i
, H(q
i
, p
i
) = q
i
p
i
L(q
i
, q
i
) . (1.2)
On quantisation, the coordinate and momenta become operators q
i
, p
i
satisfying the (equal
time) commutation relations
[ q
i
(t), p
j
(t)] = i
i
j

1. (1.3)
1
Fock, Vladimir Aleksandrovich (1898-1974)
2
Lagrange, Joseph Louis (1736-1813)
3
Hamilton, Sir William Rowan (1805-1865)
1
The generalisation of this in eld theory for scalar elds is
_

(x, t),

(y, t)

= i
(3)
(x y)

1. (1.4)
(From now on = 1).
The canonical approach is non-covariant and relativistic invariance is lost because of requir-
ing equal times (and thus picking a preferred frame in which time is measured). Lorentz
4
invariance is not manifest in a conspicuous way. That makes it hard to derive perturbation
expansions in terms of Feynman
5
rules.
Feynman invented the path integral approach, which avoids a non-covariant approach, and
is equivalent to the standard operator approach. It is better equipped for dealing with gauge
theories, for example.
1.2 Path Integral in One-Particle Quantum Mechanics
In this section the path integral for a single particle will be derived, given operators p, q and a
classical Hamiltonian
H(q, p) =
p
2
2m
+V (q) , (1.5)
which corresponds to a quantum mechanical operator

H = H( q, p) , (1.6)
where q, p satisfy [ q, p] = i

1. The Schr odinger

6
equation
i

t
[) =

H[) (1.7)
determines the time evolution of states. The (formal) solution to the
Schr odinger equation is
[(t)) = exp(i

Ht)[(0)) , (1.8)
since

H is independent of time.
We can also consider position states, satisfying
q(t)[q, t) = q[q, t) , (1.9)
where q is any real number. We use the convenient normalisation
q

, t[q, t) = (q

q) . (1.10)
Here consider the Schr odinger picture, where states [(t)) depend on time, and operators
q, p are xed. Therefore, the states [q) are time-independent, and form a basis for any state.
We can dene a wave function
(q, t) = q[(t)) , (1.11)
and acting on wave functions

H
1
2m
d
2
dq
2
+V (q) . (1.12)
The path integral approach expresses time evolution of states in terms of possible trajectories
of particles. First write the wave function as
(q, t) = q[ exp(i

Ht)[(0)) (1.13)
and introduce a complete set of states [q
0
), such that

1 =
_
dq
0
[q
0
)q
0
[ , (1.14)
4
Lorentz, Hendrik Antoon (1853-1928), Nobel Prize 1902
5
Feynman, Richard Phillips (1918-1988), Nobel Prize 1965
6
Schrodinger, Erwin Rudolf Josef Alexander (1887-1961), Nobel Prize 1933
2
to rewrite this result as
(q, t) =
_
dq
0
q[ exp(i

Ht)[q
0
)q
0
[(0)) =
_
dq
0
K(q, q
0
; t) (q
0
, 0) . (1.15)
The Schr odinger equation has been converted into an integral expression, where
K(q, q
0
; t) := q[ exp(i

Ht)[q
0
) . (1.16)
Consider the evolution of the wave function over a time interval 0 t T and divide this
interval up into smaller intervals 0 = t
0
< t
1
< t
2
< . . . < t
n+1
= T. (These intervals are not
necessarily of equal length, though it may often be convenient to choose them to be so.) Then
rewrite
exp(i

HT) = exp(i

H(t
n+1
t
n
)) exp(i

H(t
n
t
n1
)) . . . exp(i

Ht
1
) . (1.17)
At each t
r
, where r = 1, . . . , n, now introduce complete sets of states:
K(q, q
0
; T) =
_
n

r=1
_
dq
r
q
r+1
[ exp
_
i

H(t
r+1
t
r
)
_
[q
r
)
_
q
1
[ exp(i

Ht
1
)[q
0
) , (1.18)
where we set q
n+1
= q. That means that the integration goes over all possible values of q at
t
1
, t
2
, . . . , t
n
, as the value of q evolves from q
0
at t = 0 to q at t = T:
-
6
t
q
T 0 t
1
t
2
t
3
. . .

@
@
@
H
H
H
@
@
@

@
@
@
q
q
0
Up to now, the calculation has only been complicated, and there is no straightforward way for
an explicit solution of the problem for an arbitrary potential V (q).
We will approximate the factors
q
r+1
[ exp(i

Ht)[q
r
) (1.19)
arising in (1.18) for small t = t
r+1
t
r
. But rst consider a case where these expressions can
be evaluated explicitly, that is free theory, with V = 0. In fact, the result
K
0
(q, q

; t) := q[ exp(i

p
2
2m
t)[q

) (1.20)
is valid for any value of t. To show this we use a complete set of momentum states [p) (i.e.
states which satisfy p[p) = p[p)) such that

1 =
_
dp
2
[p)p[ . (1.21)
A basic result from quantum mechanics is that
q[p) = e
ipq
, (1.22)
which can be easily obtained by noting that when acting on wave functions, p i
d
dq
. Fur-
thermore,
p[q

) = e
ipq

. (1.23)
3
This means that
K
0
(q, q

; t) =
_
dp
2
q[ exp(i

p
2
2m
t)[p)p[q

) =
_
dp
2
e
i
p
2
2m
t
e
ip(qq

)
. (1.24)
Upon the substitution p

= p
m(qq

)
t
this becomes
K
0
(q, q

; t) = e
im
(qq

)
2
2t
_
dp

2
e
i
p
2
2m
t
= e
i
m(qq

)
2
2t
_
m
2it
, (1.25)
where in the last step the integrand was rotated by p

= e
i

4
r, so that
_
dp

2
e
i

2
p
2
= e
i

4
_
dr
2
e

2
r
2
=
1
2
e
i

4
_
2

. (1.26)
Note that in the limiting case t 0, we have K
0
(q, q

; t) (q q

), as necessary.
Let us return to (1.18), our integral expression for K(q, q
0
; T):
K(q, q
0
; T) =
_
n

r=1
_
dq
r
q
r+1
[ exp(i

H(t
r+1
t
r
)[q
r
)
_
q
1
[ exp(i

Ht
1
)[q
0
) .
We will in due course take the limit n and t
r+1
t
r
0 for each r. We are concerned
with nding an approximation for exp(i

Ht) for small t. One thing to note is that if

A and

B are operators, then in general

exp(

A+

B) ,= exp(

A) exp(

B) . (1.27)
In fact, one can obtain an expression
exp(

A) exp(

B) = exp
_

A+

B +
1
2
[

A,

B] +. . .
_
. (1.28)
However, we can say that for small
exp((

A+

B)) = exp(

A) exp(

B)
_
1 +O(
2
)
_
, (1.29)
since all following terms in the summation include commutators of

A and

B. This implies
exp(

A+

B) = lim
n
_
exp(

A/n) exp(

B/n)
_
n
. (1.30)
Hence, if t is small, we can write
exp(i

Ht) = exp
_
i
p
2
2m
t
_
exp
_
iV ( q)t
__
1 +O(t
2
)
_
, (1.31)
and so to leading order as t 0,
q
r+1
[ exp(i

Ht)[q
r
) = q
r+1
[ exp
_
i

p
2
2m
t
_
exp(iV ( q)t)[q
r
) . (1.32)
Because the states [q
r
) are eigenstates of the operator q, the second exponential can actually
be factored out:
q
r+1
[ exp(i

Ht)[q
r
) = e
iV (q
r
)t
q
r+1
[ exp
_
i

p
2
2m
t
_
[q
r
)
=
_
m
2it
e
i
1
2
m(
q
r+1
q
r
t
)
2
tiV (q
r
)t
, (1.33)
using the result (1.25) derived before for the free theory.
We return to the task of calculating K(q, q
0
; T) in (1.18), conveniently writing t
r+1
t
r
= t,
so that the time interval is divided into equal increments:
K(q, q
0
; T) =
_
m
2it
_1
2
(n+1)
_
n

r=1
dq
r
e
i
P
n
r=0
(
1
2
m(
q
r+1
q
r
t
)
2
V (q
r
))t
, q
n+1
= q . (1.34)
4
The path integral is obtained by considering the limit as n and t 0. Firstly in the
exponent
n

r=0
_
m
2
_
q
r+1
q
r
t
_
2
V (q
r
)
_
t S[q] , (1.35)
where
S[q] =
T
_
0
dt
_
1
2
m q
2
V (q)
_
=
T
_
0
dt L(q, q) . (1.36)
Here L(q, q) =
1
2
m q
2
V (q) is the classical Lagrangian and the notation S[q] denotes a
dependence on a function q(t) dened on the given interval. While a function assigns a number
to a number, S is a functional which assigns a number to a given function, the action. Note
that the functions q(t) have the following property
q(0) = q
0
, q(T) = q . (1.37)
As a formal denition, we also let
_
m
2it

n

r=1
__
m
2it
dq
r
_
d[q] . (1.38)
The expression d[q] can be given a more precise mathematical meaning in certain situations,
but we take the limit as a physicist, not worrying about any mathematical idiosyncrasies which
can be very non trivial. If necessary it should be dened by the limit of the discrete product
given on the left of (1.38) as n and t =
T
n
0. What then needs to be done is to show
that limit exists for any V (q). However ultimately, we can write for the limit
K(q, q
0
; T) = q[ exp(i

HT)[q
0
) =
_
d[q] e
iS[q]
, (1.39)
and usually the precise details of the limiting process may be ignored.
The path integral is a sum over all paths between q and q
0
weighted by e
iS
. It represents
the fact that the classical concept of a trajectory has no validity in quantum mechanics; in the
double-slit experiment, it is impossible to tell which path a particle has taken that is registered
on the screen. Unlike in classical physics, where there usually is a unique path, or at least
a nite number of paths, and certainly not a continuous range, we have to take all possible
trajectories into account when dealing with quantum mechanics.
This path integral (or functional integral) formalism is an alternative approach to quantum
mechanics which in many ways is quite intuitive. To illustrate it in more detail, let us consider
an example where the path integral can be carried out explicitly.
1.2.1 Example: The Harmonic Oscillator
Consider the potential
V (q) =
1
2
m
2
q
2
. (1.40)
This is the harmonic oscillator, of course. It is one of basically two solvable problems in quantum
mechanics, the second one being the hydrogen atom.
7
All paths q(t) have the property that
q(0) = q
i
, q(T) = q
f
. (1.41)
In this case, the classical path will play a role in the evaluation of the path integral. We will
expand q(t) about the classical path q
c
(t), which is dened by obeying
q
c
+
2
q
c
= 0, q
c
(0) = q
i
, q
c
(T) = q
f
. (1.42)
7
The path integral for the hydrogen atom was famously solved by Kleinert, Hagen of course (developing
work also with I.H. Duru).
5
So can we solve this dierential equation? The best way to write the solution is the one which
is the simplest to write down. All solutions include sint and cos t, and one could write the
solution as a linear combination of these and then work out the prefactors. However it is much
nicer to choose instead sint and sin(T t).
8
You can then write down the answer almost
by inspection:
q
c
(t) =
1
sinT
_
q
f
sint +q
i
sin(T t)
_
. (1.43)
Clearly this satises the dierential equation, and it is easy to verify that the boundary condi-
tions are also satised, so we have solved (1.42).
The action for the particular path is given by
S[q
c
] =
1
2
m
T
_
0
dt
_
q
2
c

2
q
2
c
_
. (1.44)
Integrate this by parts
S[q
c
] =
1
2
m
_
q
c
q
c

T
0

1
2
m
T
_
0
dt q
c
( q
c
+
2
q
c
) , (1.45)
where the last integral vanishes since q
c
obeys the equation of motion (1.42). We know what
values q
c
itself takes at t = 0 and t = T:
S[q
c
] =
1
2
m
_
q
f
q
c
(T) q
i
q
c
(0)
_
. (1.46)
Now use the explicit solution for q
c
to obtain
S[q
c
] = m
2q
f
q
i
+ (q
f
2
+q
i
2
) cos T
2 sinT
. (1.47)
Note as a consistency check that as 0,
S[q
c
]
1
2
m
(q
f
q
i
)
2
T
. (1.48)
We can write a general path as
q(t) = q
c
(t) +f(t), f(0) = f(T) = 0 . (1.49)
For S[q] quadratic in q then
S[q] = S[q
c
] +S[f] (1.50)
is exact since as S[q] is stationary at q = q
c
. (If you vary the action, you get the classical
equations of motion - that is what the action came from.) That is, there is no linear term
in f. This is analogous to expanding a function around a minimum, where the rst terms in
the expansion will be the value of the function at the minimum and a term quadratic in the
deviation from the minimum.
We have also assumed that f is small, meaning that only paths close to the classical path
will contribute to the integral. Furthermore, we can assume
d[q] = d[f] , (1.51)
which is something which is true for ordinary integrals, namely that d(x+a) = dx for a constant
a. (This also makes sense in terms of the more formal denition (1.38) which we gave.) We can
now write that in this particular example,
K(q, q
0
; T) =
_
d[q] e
iS[q]
= e
iS[q
c
]
_
d[f]e
iS[f]
. (1.52)
8
These are actually identical in the case T = n for some integer n. In this case, there is no classical
solution unless q = (1)
n
q
0
. This special case is ignored here.
6
Note that the integral is independent of q
0
and q, thus all the dependence on the initial and
nal points is contained in the prefactor. There are various ways to derive the second factor;
we use one which is potentially useful later on.
Expand f(t) in terms of a convenient complete set. We use a Fourier sine series for f:
f(t) =

n=1
a
n
_
2
T
sin
nt
T
. (1.53)
The sine functions form an orthonormal basis for functions vanishing at t = 0 and t = T. We
can write, integrating by parts, noting f(0) = f(T) = 0 and using orthonormality,
S[f] =
1
2
m
T
_
0
dt f(

f +
2
f) =
1
2
m

n
a
2
n
_
n
2

2
T
2

2
_
. (1.54)
We assume the relation
d[f] = C

n=1
da
n
, (1.55)
where C is a normalisation constant.
9
We are now in the position to express the integral in the
form
_
d[f] e
iS[f]
= C

n=1
_
da
n
e
i
m
2
a
2
n
(
n
2

2
T
2

2
)
. (1.56)
The basic integral (solved by rotating the contour) here is

dy e
i
2
y
2
=
_
2i

. (1.57)
It is convenient to write now
_
d[f] e
iS[f]
= C
0

n=1
1
_
1

2
T
2
n
2

2
, (1.58)
where we have absorbed constant factors like (

n=1
n)
1
into C
0
. This factor is divergent, but
does not depend on any of the critical parameters. (We will not talk about innities appearing
here.) In the free case = 0, the product is equal to one and we are left with
C
0
=
_
m
2iT
, (1.59)
which xes the normalisation.
What can we say about this innite product? It can be shown that

n=1
_
1

2
T
2
n
2

2
_
=
sinT
T
. (1.60)
(Note that both sides have the same zeros as functions of T. Furthermore, they both go to
one as T 0. Several other observations show that both sides have the same behaviour and
actually are identical.)
Ultimately,
_
d[f] e
iS[f]
=
_
m
2i sinT
, (1.61)
9
Thinking about the formal denition (1.38), if we consider t
r
= r, with = T/(n + 1), then the trans-
formation from {f(t
r
) : r = 1, . . . , N} to {a
n
: n = 1, . . . , N} is an orthogonal transformation so that
Q
r
df(t
r
) =
Q
n
da
n
.
7
which is a nice everyday function. The overall result, in all its glory, is the following:
K(q
f
, q
i
; T) =
_
m
2i sinT
e
im
(q
f
2
+q
i
2
) cos T2q
f
q
i
2 sin T
. (1.62)
To check this, note that there are eigenfunctions [n) of

H with
E
n
=
_
n +
1
2
_
,

1 =

n
[n)n[ . (1.63)
A formula which can be obtained by using standard quantum mechanics is
K(q
f
, q
i
; T) =

n
(q
f
)

n
(q
i
)e
i(n+
1
2
)T
. (1.64)
In the situation where T = i, ,
2i sinT e

, 2 cos T e

. (1.65)
We nd that
K(q
f
, q
i
; i)

_
m

1
2

1
2
m(q
f
2
+q
i
2
)
=
0
(q)

0
(q
0
)e

1
2

. (1.66)
In this special case, the result obtained from the path integral calculation is consistent with
standard quantum mechanics.
1.2.2 A Few Comments
(i) Generally speaking, integrals of the form
_
dx e
ix
2
are rather ill-dened because they do
not converge (absolutely). It is much better to consider integrals
_
dx e
x
2
for > 0.
The same thing happens with path integrals; note that
S[q] =
T
_
0
dt
_
1
2
m q
2
V (q)
_
(1.67)
is normally a real quantity. But in order to obtain well-dened integrals, we can consider
an analytic continuation of time
t i, T i
1
, (1.68)
so that
q
2

_
dq
d
_
2
, iS[q]

1
_
0
d
_
1
2
m
_
dq
d
_
2
+V (q)
_
(1.69)
and the path integral becomes
q[ exp
_

H
_
[q
0
) =
_
d[q] e

1
0
d (
1
2
m(
dq
d
)
2
+V (q))
. (1.70)
The point about this is that the right-hand side integral has a rigorous denition for a
wide class of potentials V , subject to the requirement that V is bounded from below.
In many contexts, one considers these integrals with analytic continuation. They rst
appeared in the context of Brownian
10
motion.
(ii) The path integral provides a method of making non-perturbative approximations; we will
show this for the example of tunnelling. A widely known result from quantum mechanics
is that particles can tunnel through a potential barrier. Consider a situation where we
have a potential
10
Brown, Robert (1773-1858)
8
-
6
V (q)
q q
0
q
1
We want to calculate the amplitude to get from q
0
at time t
0
to q
1
at time t
1
, eventually
taking the limits t
0
and t
1
. So use the path integral to evaluate
q
1
[ exp
_
i

H(t
1
t
0
)
_
[q
0
) . (1.71)
One way of proceeding with these path integrals is to expand around a classical path
q(t) = q
c
(t) +f(t) , (1.72)
where the classical path q
c
(t) satises the classical equations and given boundary condi-
tions. This method was used in the example above. In general, there will not necessarily
be a classical path; here this is in the case when the total energy is smaller than the max-
imum of the potential between q
0
and q
1
. We make use of analytic continuation t i,
such that
iS[q] =

1
_

0
d
_
1
2
m
_
dq
d
_
2
+V (q)
_
, (1.73)
and we are interested in the limit
0
,
1
. (Note that this actually means we
choose a dierent contour in the complex plane to evaluate the integral. This is a method
commonly used to evaluate integrals over analytic functions, and we have in fact already
used it above. One makes use of this in the method of steepest descents, for example.)
The classical equation for q
c
() is now
m
d
2
d
2
q
c
+V

(q
c
) = 0 , (1.74)
which we integrate once to get

1
2
m
_
dq
c
d
_
2
+V (q
c
) = E . (1.75)
This is similar to classical mechanics, but with one sign ipped because of our funny
change in time. We want a situation in which as , q() q
0
or q() q
1
, where
V (q
0
) = V (q
1
) = 0. But if it is smoothly going to these points we must also have
dq
d

q=q
0
=
dq
d

q=q
1
= 0 E = 0 . (1.76)
In this situation, we can actually solve this, assuming q
1
> q
0
and therefore taking the
positive square root to get
dq
c
d
=
_
2V (q
c
)
m
. (1.77)
Now substitute this in to evaluate iS[q
c
]:
iS[q
c
] = 2

d V (q
c
) =
q
1
_
q
0
dq
_
2mV (q) , (1.78)
because d =
dq

m

2V (q)
for this solution.
9
As we have seen in the case of the harmonic oscillator, the dependence of the path integral
on the initial and nal points is contained in e
iS[q
c
]
, so the tunelling amplitude will be
proportional to
E

R
q
1
q
0
dq

2mV (q)
. (1.79)
This is an exponential suppression, which is a real quantity, but non-zero in all cases. (It
is known as the Gamow
11
factor.)
The path integral calculation gives the same result as WKB, for example, in a relatively
simple way.
(iii) In general, there is no unique correspondence
H(q, p)

H = H( q, p) (1.80)
in quantum mechanics, because a product pq can be replaced by
pq
_
p q
q p
, (1.81)
and these are not equal (since [ q, p] ,= 0). In many problems there are ways to resolve
this ambiguity, but it is still there. This is reected in the path integral, although this
is far from obvious. There are more complicated problems where dierent discretisations
of the path integral will give dierent answers. Although we will not be very concerned
with this problem, it is worth bearing in mind it exists.
In due course, we are going to apply the ideas of path integrals to eld theories. But rst
let us do some calculations which will be useful later on.
1.3 Gaussian Integrals and Extensions Over Multi-Dimensional Co-
ordinates
Consider a set of coordinates, represented by a column vector
x = (x
1
, . . . , x
n
)
T
R
n
, (1.82)
and dene a scalar product
x x

= x
T
x

=
n

i=1
x
i
x

i
. (1.83)
Let A be a n n symmetric positive denite matrix and consider the Gaussian
12
integral
Z
A
=
_
d
n
x e

1
2
xAx
(1.84)
This kind of integral is essentially equivalent to free eld theory, as will become apparent in
due course.
Its evaluation is quite simple: note that there is an orthogonal matrix U ([ det U[ = 1) such
that
U AU
T
= D =
_
_
_

1
0
.
.
.
0
n
_
_
_ , (1.85)
where
i
> 0 for all i = 1, . . . , n. It is now quite straightforward to compute the integral with
a suitable change of variable; namely let
x

= U x, d
n
x = d
n
x

(1.86)
11
Gamow, George (1904-1968)
12
Gau, Carl Friedrich (1777-1855)
10
Z
A
=
_
d
n
x

1
2
x

Dx

=
n

i=1
_
dx

i
e

1
2

i
x

i
2
, (1.87)
so the integral factorises because of
x

Dx

=
n

i=1

i
x

i
2
. (1.88)
The generic integral is of the form
_
dx e

1
2
x
2
=
_
2

. (1.89)
So now we know the answer
Z
A
=
n

i=1
_
2

i
=
(2)
n
2

det A
. (1.90)
An important generalisations is to the complex case, when we have a column vector
z = (z
1
, . . . , z
n
)
T
C
n
, (1.91)
and dene a scalar product
z z

= z

=
n

i=1
z

i
. (1.92)
In general, when z = x +iy, we dene
d
2n
z = d
n
xd
n
y . (1.93)
Let us take B to be a Hermitian
13
matrix with positive (real) eigenvalues and consider the
Gaussian integral
Z
B
=
_
d
2n
z e
zBz
. (1.94)
In this case, there is a unitary matrix U such that
U BU

= D =
_
_
_

1
0
.
.
.
0
n
_
_
_ , (1.95)
where again all
i
> 0. Now the same trick as before applies:
z

= U z , d
2n
z = d
2n
z

, (1.96)
and
z

Dz

=
n

i=1

i
[z

i
[
2
. (1.97)
The basic integral here is
_
d
2
z e
|z|
2
=
_
dx dy e
(x
2
+y
2
)
=

. (1.98)
(Note that
_
dz usually denotes line integrals, where here we integrate over the whole complex
plane.) The result in this case is
Z
B
=

n
det B
, (1.99)
which should be compared to (1.90) - note that in this case the determinant of the matrix
appears instead of its square root.
13
Hermite, Charles (1822-1901)
11
As a brief comment, let us consider an i in the exponent
Z
A
=
_
d
n
x e

i
2
xAx
=
_
2
i
_n
2
(det A)

1
2
, (1.100)
where we have analytically continued each x integral into the complex plane.
Now consider the extension to a linear term in the exponent
Z
A,b
=
_
d
n
x e

1
2
x Ax+bx
. (1.101)
We can reduce this one to the previous case in the following way. Note that we can write
1
2
x Ax b x =
1
2
x

Ax

1
2
b A
1
b , (1.102)
where
x

= x A
1
b . (1.103)
(Since the eigenvalues of A are assumed to be positive, there is no problem whatsoever in
dening the inverse.) With the result (1.90) obtained above, we then get
Z
A,b
= e
1
2
bA
1
b

_
d
n
x

1
2
x

Ax

= e
1
2
bA
1
b

(2)
n
2

det A
. (1.104)
The analogous formula in the complex case is
Z
B,b
=
_
d
2n
z e
zBz+

bz+ zb
= e

bB
1
b
Z
B
. (1.105)
Next we will consider how integrals can be expanded and how these expansions can be repre-
sented by pictures. This will lead to Feynman graphs.
We will discuss things which are called expectation values:
f(x)) =
1
Z
A,0
_
d
n
x f(x)e

1
2
xAx
. (1.106)
In general we are interested in functions which are polynomial in some sense. We will usually
assume that all relevant functions may be expanded in a power series, so that we can restrict
our considerations to polynomial functions (by linearity). Note that by denition,
1) = 1 . (1.107)
The major trick we are going to use here is the observation that
f(x) = f
_

b
_
e
bx

b=0
, (1.108)
where

b

b
=
_

b
1
, . . . ,

b
n
_
. (1.109)
The observation then follows directly from

b
e
bx
= xe
bx
. (1.110)
It follows therefore that we have a trick for evaluating the expectation value f(x)). We can
write
f(x)) =
1
Z
A,0
_
d
n
x f
_

b
_
e

1
2
xAx+bx

b=0
, (1.111)
and take f outside the integral:
f(x)) =
1
Z
A,0
f
_

b
_
Z
A,b

b=0
. (1.112)
12
We can simplify this straight away by substituting in our previous result (1.104)
f(x)) = f
_

b
_
e
1
2
bA
1
b

b=0
. (1.113)
This is a very convenient way for working out these expectations.
Let us consider some particular cases:
x
i
) = 0 , (1.114)
x
i
x
j
) =

b
i
(A
1
b)
j
e
1
2
bA
1
b

b=0
= A
1
ij
. (1.115)
We introduce the notation that a component of the inverse of the matrix A is represented by a
line, whose ends are labelled by i and j:
i j
A little calculation will show that
x
i
1
x
i
2
. . . x
i
n
) = 0 for odd n, (1.116)
and also
x
i
x
j
x
k
x
l
) = A
1
ij
A
1
kl
+A
1
ik
A
1
jl
+A
1
il
A
1
jk
. (1.117)
(Note that A
1
ij
denotes the component ij of the matrix A
1
, and not the inverse of the
component ij of the matrix A, A
1
is symmetric since A is.)
Draw a picture for this case:
i j
k l
i j
k l

i j
@
@
@
k l
We can give a diagrammatic picture of how these things work.
The second result can be obtained in a slightly dierent way: Consider
A
ik
x
k
x
j
) =
1
Z
A,0
_
d
n
x A
ik
x
k
x
j
e

1
2
xAx
=
1
Z
A,0
_
d
n
x x
j

x
i
_
e

1
2
xAx
_
. (1.118)
Integration by parts gives
A
ik
x
k
x
j
) =
1
Z
A,0
_
d
n
x
_

x
i
x
j
_
e

1
2
xAx
=
ij
. (1.119)
So the matrix A
ik
acting on x
k
x
j
) gives the identity, which means that (because of symmetry)
x
i
x
j
) = A
1
ij
. (1.120)
The complex case contains no essential new ideas as compared to the real case, and is
summarised here: Let B be Hermitian, then
f(z, z)) :=
1
Z
B,0
_
d
2n
z f(z, z) e
zBz
=
1
Z
B,0
f
_

b
,

b
_
Z
B,b

b=0
. (1.121)
One nds that
z
i
z
j
) = B
1
ij
, (1.122)
which is no longer necessarily symmetric in i and j. Therefore we denote this by
13
- i j
i.e. we add an arrow to denote which index comes rst.
Exercise: Obtain this result by showing that
B
ik
z
k
z
j
) =
ij
= z
i
z
k
)B
kj
. (1.123)
Generally, to visualise x
i
1
. . . x
i
2n
), we draw n lines linking dierent points. One can write
down diagrammatically what an expectation value is by drawing more and more lines.
1.4 Non-Gaussian Integrals and Perturbation Expansions
These integrals will correspond to interacting eld theory (whereas the previous Gaussian
integrals corresponded to free eld theory). The basic integral in this case is
Z =
1
Z
A,0
_
d
n
x e

1
2
xAx+bxV (x)
, (1.124)
where the prefactor conveniently ensures that Z

b=0,V =0
= 1. We require that the real function
V (x) is bounded below, and also V (0) = 0. Now we use the same trick as before:
Z = e
V (

b
)
1
Z
A,0
_
d
n
x e

1
2
xAx+bx
. (1.125)
This is very formal, but for nite integrals there is no real problem in writing down this expres-
sion. Substituting in, we get
Z = e
V (

b
)
e
1
2
bA
1
b
. (1.126)
We can evaluate this by expanding e
V (

b
)
. This will give the rise to a perturbation expansion.
However, there is a slightly alternative way of doing it which from some point of view makes
the manipulations a little easier. V may be a complicated function.
Lemma
G
_

b
_
F(b) = F
_

x
_
G(x)e
xb

x=0
. (1.127)
Proof
We will show this for a special case: Let G(x) = e
x
, F(b) = e
b
. Then the left-hand side is
G
_

b
_
F(b) = e

F(b) = F(b +) = e
(b+)
. (1.128)
Let us attempt to consider the right-hand side:
f
_

x
_
G(x)e
xb

x=0
= e

x
e
x(+b)

x=0
= e
(x+)(+b)

x=0
= e
(+b)
. (1.129)
The result is then true for any F and G as one may express F and G as a Fourier series.
Let us apply this to the expression for Z, so that
Z = e
1
2

x
A
1
x
e
V (x)+bx

x=0
. (1.130)
We get a perturbative expansion by expanding both exponentials, setting for simplicity b = 0.
We use the notation
V
i
1
i
2
...i
k
=

x
i
1

x
i
2
. . .

x
i
k
V (x)

x=0
, (1.131)
where i
1
, . . . , i
k
1, . . . , n. The rst few terms of the expansion are, assuming that V (0) =
V
i
(0) = 0,
Z = 1
1
2
A
1
ij
V
ij

1
8
A
1
ij
A
1
kl
V
ijkl
+
1
8
A
1
ij
V
ij
A
1
kl
V
kl
+
1
4
V
ij
A
1
ik
A
1
jl
V
kl
+
1
8
V
ijk
A
1
ij
A
1
kl
A
1
mn
V
lmn
+
1
12
V
ijk
A
1
il
A
1
jm
A
1
kn
V
lmn
+. . . . (1.132)
14
These expressions are not very transparent; proliferation of indices makes it hard to see what
is going on.
Diagrammatic interpretation:
We represent A
1
ij
by a line joining the points i and j, as before, and V
i
1
i
2
...i
k
by a vertex
joining k lines, e.g.

A
A
A
A

i
1
i
2
i
3
i
4
i
5
i
6
Then we can associate the terms shown above in the expansion of Z 1 with the diagrams

1
2
`
_

1
8
`
_
`
_
+
1
8
`
_
`
_
+
1
4
`
_
+
1
8
`
_
`
_
+
1
12
`
_
+. . .
Diagrams like the third one are called disconnected. In general, all these diagrams are called
vacuum diagrams, since there are no external lines. Check the following, known as Eulers
14
formula, holds for all connected diagrams:
L = I V + 1 , (1.133)
where L is the number of closed loops
15
, I is the number of internal lines and V is the number
of vertices. To demonstrate consider the subgraph where the V vertices are linked by just V 1
lines to create a minimal connected tree graph (this may of course not be unique). Then add
the remaining I V +1 lines between the various vertices to restore the complete graph; each
addition creates a new loop.
Let us summarise the Feynman rules for diagrammatic expressions of integrals of this type:
Lines, with end points labelled by i and j, represent A
1
ij
.
Vertices represent V
i
1
...i
k
.
Contract all indices.
Now take the general case b ,= 0; let us see how this is modied. We can now always maintain
V
i
(0) = 0 for all i by redening V (x), absorbing all contributions into b. We introduce external
lines
, b
i
associated with one b
i
. The rst terms in the expansion which will now also contribute are
1
2
b
i
A
1
ij
b
j

1
6
b
i
b
j
b
k
A
1
il
A
1
jm
A
1
kn
V
lmn

1
2
b
i
b
j
A
1
ik
A
1
jl
V
kl
+. . . . (1.134)
1
2
, ,

1
6
P
P
P

,
,
, ,

1
2
, , ,
14
Euler, Leonhard (1707-1783)
15
For a graph which can be drawn in the plane, this means the number of faces, or regions separated by
edges. In general, the denition is slightly harder to see geometrically. The most natural way to think of this
for our purposes is as follows: imagine some conserved number (like momentum) is owing through the graph;
then if you are told how much goes in/out at the external points, you can deduce something about ow within
the graph. However, an arbitrary amount can ow around a circle in the graph. The number of independent
such amounts we will later think of as the number of loops.
15
In all of these expressions, all lines are attached to external lines, there are no superuous
indices. Of course, the complexity increases quite dramatically as one proceeds. Diagrams such
as those just described which have no loops are called tree diagrams vacuum diagrams always
have loops). A one loop diagram with two external lines is
. . . +
1
4
b
i
b
j
A
1
ik
A
1
jl
V
kmn
V
lpq
A
1
mp
A
1
nq
+. . . . (1.135)
1
4
, , , ,
For each diagram it is necessary to also include an associated coecient
1
S
, where S is termed
the symmetry factor of the diagram. For any graph there is a symmetry or automorphism
group o

which is the group of permutations of lines and vertices which leave the graph invariant.
A graph consists of vertices v and lines l linking vertices, for o

we require that
for all l linking v to v

then (l) links (v) to (v

) or, if the lines have no direction, (v

)
to (v). S is the order of o

, S = [o

[, so that it is the number of ways lines and vertices

may be permuted leaving the diagram invariant. External lines with labels such a momenta
cannot be permuted as this would not leave the graph invariant. If the graph has only vertices
linked by single lines then the symmetry group is found by considering just permutations of the
vertices which leave it invariant. For the graphs representing the expansion of Z it may also
be dened as the group formed by non trivial permutations of the indices associated with the
vertices V
ij
= V
(ij)
, V
ijk
= V
(ijk)
, . . . which leave the expression corresponding to the diagram
unchanged (trivial permutations are those which are just a dierent relabelling of all the indices,
which are just summed over dummy variables). For instance, the second diagram in the sum
above has S = 3! = 6, since all three lines can be commuted. The last diagram has S = 4,
since both external lines can be commuted, as well as the lines in the middle. It takes some
experience to work these factors out in practice, and it is always worth trying to check the
results, for it is not very dicult to get these things wrong.
To actually prove that the coecient is indeed
1
S
is less straightforward. Suppose the graph
is a collection p
n
n-vertices, at which n identical lines are incident. External lines can be
regarded as vertices with n = 1. Each vertex has an associated 1/n! from the expansion of V
and also 1/p
n
! from the expansion of the exponential. Label the graph uniquely by giving all
lines a unique label. If we count the number of ways of forming a particular labelled graph
then there is an n! for each n-vertex, from the n! ways of coupling the labelled lines to each
n-vertex, and also a p
n
! from permuting all the n-vertices. These factors then cancel the 1/n!
and 1/p
n
! factors. However the Feynman integrals correspond to unlabelled graphs, each of
which generates S

unique labelled graphs so it is necessary to introduce a factor 1/S which

may be regarded as arising from an incomplete cancellation of the 1/n! and 1/p
n
! factors.
We now consider a couple of special cases, just to verify how this works:
Let b = 0 and V
i
1
...i
k
= 0 for k 3. If we represent the second derivatives of V by an (nn)
matrix V

, the result you can write is

Z = 1
1
2
tr(A
1
V

) +
1
8
_
tr(A
1
V

)
_
2
+
1
4
tr(A
1
V

A
1
V

) +. . . , (1.136)
where tr(A
1
V

) = A
1
ij
V

ji
, of course. But the problem can also be solved exactly, with the
exact result being given by
Z =
1
Z
A
_
d
n
x e

1
2
x(A+V

)x
=
1
Z
A
(2)
n
2
_
det(A+V

)
_

1
2
(1.137)
=
_
det A
1
det(A+V

)
_

1
2
=
_
det(I +A
1
V

)
_

1
2
, (1.138)
where (det M)
1
= det(M
1
) and det(M N) = det M det N . was used.
For any (diagonalisable) matrix M, we have the following identity
log det M = tr log M , (1.139)
16
which follows from the observation that the determinant of a matrix is the product of its
eigenvalues and a logarithm of a product is the sum of logarithms. For an eigenvalue
i
= 0,
the identity becomes singular.
We use this identity and the expansion log(1 +x) = x
x
2
2
+. . . to obtain an expansion for
Z:
Z = e

1
2
log det(I+A
1
V

)
= e

1
2
tr log(I+A
1
V

)
= e

1
2
tr(A
1
V

+
1
4
(A
1
V

A
1
V

)+...)
= 1
1
2
tr(A
1
V

) +
1
8
_
tr(A
1
V

)
_
2
+
1
4
tr(A
1
V

A
1
V

) +. . . , (1.140)
so we have reproduced the previous result.
Note that we can also consider b ,= 0, where only the second derivatives of V are non-zero (as
before). We will get additional terms (only second order in b)
1
2
, ,
+
1
2
, , ,
+
1
2
, , , , . . .
corresponding to
1
2
b
i
(A
1
ij
A
1
ik
V

lk
A
1
lj
+. . .)b
j
=
1
2
b
i
(A+V

)
1
ij
b
j
, (1.141)
where the expansion arises from a geometric series and the nal line is an exact result. The
series must reproduce what we would expect; recall (1.104)
Z
A,b
= e
1
2
bA
1
b
(2)
n
2

det A
.
To draw a connection to eld theory, we need to consider the following integral with a
complex exponent:
_
d
n
x e
i(
1
2
x
i
A
ij
x
j
+V (x))
. (1.142)
In this case the Feynman rules are slightly modied, namely that
Lines, with end points labelled by i and j, represent iA
1
ij
.
Vertices represent iV
i
1
...i
k
.
The symmetry factors are unchanged.
Feynman diagrams provide a very convenient shorthand for expressing expansions of inte-
grals of the exponential form which denes Z. The expansion makes sense only when V is small,
of course, and in general gives only an asymptotic series, the actual radius of convergence is
zero.
Up to this point we have played around with toy problems, path integrals in standard
quantum mechanics and calculated some integrals. The motivation for this is to extend all
these ideas to quantum elds, which is what we will now do.
17
2 Functional Methods in Quantum Field Theory
2.1 Free Scalar Field Theory
We usually have a scalar eld
(x) = (x, t) . (2.1)
Our convention for the Minkowski
16
metric is

= diag (1, +1, . . . , +1). (2.2)

The Lagrangian density for free eld theory is given by
L
0
=
1
2

1
2
m
2

2
=
1
2

1
2
(

)
2

1
2
m
2

2
, (2.3)
which leads to the action
S
0
[] =
_
d
d
x L
0
=
_
d
d
x
_
1
2
(x)(x)
_
, (2.4)
by integration by parts, where
+m
2
=

2
t
2

2
+m
2
, (2.5)
is the Klein-Gordon operator. Note that we take the number of dimensions to be d, which
can generally take any value. The classical equation of motion is
= 0, (2.6)
the Klein-Gordon
17
equation. It was actually discovered by Schr odinger, but because of
problems arising in standard quantum mechanics he then tried to nd a dierential equation
which was rst order in time. However, it is perfectly o.k. when you move to quantum eld
theory.
We can dene a quantum eld theory, instead of by using the classical approach of going
to the Hamiltonian formalism and then imposing certain commutation relations for elds and
conjugate momenta, through a functional integral:
Z
0
[J] =
_
d[] e
iS
0
[]+i
R
d
d
x J(x)(x)
. (2.7)
Square brackets, as usually, denote that Z
0
is a functional, depending on a function J(x).
J is sometimes referred to as a source, like an external current in classical electrodynamics.
Formally,
d[] =

x
d(x) . (2.8)
We sort of dene it in the following way

x
d(x) lim
a0

i
d(x
i
) , (2.9)
where the points x
i
belong to a lattice of size a. One will rst consider a nite volume for this
lattice, then take the limit of innite volume. This is the same approach as in one-dimensional
quantum mechanics, where we divided a time interval into smaller intervals, taking the limit of
the size of the steps going to zero.
We introduce the idea of a functional derivative, dened by the essential rule
J(y)
J(x)
=
d
(x y) , (2.10)
16
Minkowski, Hermann (1864-1909)
17
Klein, Oskar (1894-1977), Gordon, Walter (1893-1939)
18
together with the usual Leibniz and chain rules for dierentiation.
This means that for instance,

J(x)
e
i
R
d
d
x J(x)(x)
= i(x)e
i
R
d
d
x J(x)(x)
. (2.11)
Note that

(x)
S
0
[] = (x) , (2.12)
since S
0
[] =
_
d
d
x (x)(x).
In order to evaluate the functional integral, we replace
S
0
[] +
_
d
d
x J(x)(x) = S
0
[

] +
1
2
_
d
d
x J(x)
1
J(x) , (2.13)
where

(x) = (x)
1
J(x); note that then
S
0
[

] =
1
2
_
d
d
x
_
(x)(x) (x)J(x) +J(x)
1
J(x)
1
J(x)(x)
_
=
1
2
_
d
d
x
_
(x)(x) 2(x)J(x) +J(x)
1
J(x)
_
. (2.14)
If we choose the normalisation
Z
0
[0] =
_
d[] e
iS
0
[]
= 1 , (2.15)
we would then have
Z
0
[J] = e
1
2
i
R
d
d
x J(x)
1
J(x)
. (2.16)
In principle we have worked out the functional integral, but we need to nd
1
. is a
dierential operator, hence
1
is a Green
18
s function. So the next step will be to obtain
the inverse of .
Note that our nal result is the innite-dimensional analogue of the result we derived before,
Z
A,b
= Z
A
e
1
2
bA
1
b
, (2.17)
where A
1
was the inverse of the matrix A. In this case, we essentially have to solve the
equation

F
(x y) =
d
(x y) , (2.18)
so that
F
(x) is a Greens function of the operator . We can solve this fairly easily by using
Fourier
19
transformations: We dene

F
(p) =
_
d
d
x e
ipx

F
(x) =
_
d
d
x e
ip(xy)

F
(x y) , (2.19)
where p x = p

. Multiplying the above equation by e

ip(xy)
and integrating over all x
components gives

_
d
d
x e
ip(xy)

F
(x y) =
_
d
d
x
d
(x y) e
ip(xy)
= 1 . (2.20)
Since
x
=
x
+ m
2
, we can integrate the left-hand side twice by parts, dropping surface
terms to obtain

_
d
d
x
x
_
e
ip(xy)
_

F
(x y) =
_
d
d
x (p
2
+m
2
)e
ip(xy)

F
(x y) = 1 . (2.21)
18
Green, George (1793-1841)
19
Fourier, Joseph (1768-1830)
19
So the equation that you then get is
(p
2
+m
2
)

F
(p) = 1 . (2.22)
There is an ambiguity here; there is no unique solution to the above dierential equation since
one can always add a solution of the homogeneous equation to
F
(x y). We need boundary
conditions to make
F
unique. The choice

F
(p) =
1
p
2
+m
2
i
, (2.23)
denes the Feynman propagator, where > 0 guarantees that the denominator will not be
zero, and is essentially innitesimal.
Why is this an appropriate description? To motivate this, let us go back to the original
integral
Z
0
[J] =
_
d[] e

i
2
R
d
d
x (

(x)

(x)+m
2

2
(x))+i
R
d
d
x J(x)(x)
. (2.24)
This is an integral with a highly oscillating integrand; if we replace m
2
m
2
i for positive
small this will give a factor
E

1
2

R
d
d
x
2
(x)
, (2.25)
ensuring convergence of the functional integral. The integration is damped for large .
Let us try and analyse
F
(x). We can do this by going back to the Fourier transform
i
F
(x) = i
_
d
d
p
(2)
d
e
ipx
1
p
2
+m
2
i
. (2.26)
Dening E
p
=
_
m
2
+ p
2
, we can rewrite
p
2
+m
2
i = (p
0
)
2
+ (E
p
i)
2
. (2.27)
Note that this means a re-denition of , since there will now be a term of 2iE
p
on the
right-hand side (and we ignore
2
). But since the exact magnitude of does not matter and
we will in both cases get a small negative imaginary quantity appearing on both sides, this is
a valid replacement. We obtain the expression
i
F
(x) = i
_
d
d1
p
(2)
d1
_
dp
0
2
e
ip
0
t+i px
1
(p
0
)
2
(E
p
i)
2
= i
_
d
d1
p
(2)
d1
_
dp
0
2
e
ip
0
t+i px
1
2E
p
_
1
p
0
E
p
+i

1
p
0
+E
p
i
_
. (2.28)
To evaluate the p
0
integral, close the contour in the complex plane in the upper half or lower
half plane such that e
ip
0
t
0. Note that this means that for t < 0 the contour is closed in the
upper half plane, whereas for t > 0 the contour is closed in the lower half plane (which gives
an extra minus sign in the residue because of clockwise orientation):
i
F
(x) = i
_
d
d1
p
(2)
d1
1
2
1
2E
p
e
i px
_
2i(t)e
iE
p
t
2i(t)e
iE
p
t
_
, (2.29)
where (t) =
_
1, t 0
0, t < 0
is a step function. Our nal result is
i
F
(x) =
_
d
d1
p
(2)
d1
1
2E
p
e
i px
_
(t)e
iE
p
t
+(t)e
iE
p
t
_
. (2.30)
The rst term corresponds to positive frequency or positive energy particles going forward
in time, the second term corresponds to negative frequency anti-particles (or particles going
backwards in time).
20
We now have the nal expression for the generating functional
Z
0
[J] = e

1
2
R
d
d
x d
d
y J(x)i
F
(xy)J(y)
. (2.31)
We dene the two-point correlation function to be given by
(x
1
)(x
2
)) = (i)
2

J(x
1
)

J(x
2
)
Z
0
[J]

J=0
, (2.32)
a relation which is true for free elds. Explicitly this gives, since
F
(x y) =
F
(y x),
(x
1
)(x
2
)) =

J(x
1
)
_
d
d
x J(x)i
F
(x x
2
)

J=0
= i
F
(x
1
x
2
) . (2.33)
We interpret this diagrammatically as a line joining two points:
x y
Similarly we dene the n-point correlation function
(x
1
) . . . (x
n
))
(n)
= (i)
n

J(x
1
)
. . .

J(x
n
)
Z
0
[J]

J=0
. (2.34)
This is zero for any odd value of n. We can again represent the dierent contributions by
diagrams, e.g. for n = 4 there will be three contributions:
x
1
x
2
x
3
x
4
x
1
x
2
x
3
x
4

x
1
x
2
@
@
@
x
3
x
4
For free elds, we have the relation
(x
1
) . . . (x
n
)) = 0[T
_

(x
1
) . . .

(x
n
)
_
[0) , (2.35)
where T denotes time ordering and

is the eld operator. See the result in the Quantum
Field Theory course
0[T

(x)

(y)[0) = i
F
(x y) , (2.36)
which can be derived using the standard expansion in terms of creation operators a

p
and
annihilation operators a
p
.
Free elds, however, are not terribly interesting; we extend these ideas to interacting theory.
2.2 Interacting Scalar Field Theory
We include a potential V () into the action
S[] =
_
d
d
x (L
0
V ()) , (2.37)
where L
0
is the free theory Lagrangian, quadratic in the elds, and we assume the potential
to include terms of order
3
and higher.
Consider what happens when we dene the functional integral
Z[J] =
_
d[] e
iS[]+i
R
d
d
x J(x)(x)
. (2.38)
In due course we will dierentiate this with respect to J and dene correlation functions. We can
dene this integral by a perturbation expansion. This can be expressed in terms of Feynman
21
diagrams, and for each diagram there is an amplitude given by the Feynman rules. Feynman
diagrams are a pictorial way of expressing a perturbative expansion of integrals like this.
Formally, by the same tricks as previously,
Z[J] = e
i
R
d
d
x V (i

J(x)
)
Z
0
[J]
= e
i
R
d
d
x V (i

J(x)
)
e

i
2
R
d
d
xd
d
y J(x)
F
(xy)J(y)
= e
i
2
R
d
d
xd
d
y

(x)

F
(xy)

(y)
e
i
R
d
d
x (V ((x))+J(x)(x))

=0
, (2.39)
where in the rst line we have used that i

J(x)
e
i
R
J
= (x) e
i
R
J
, and from the second
to the third line we have used the innite-dimensional form the lemma derived in Section 1,
namely
G
_
i

J
_
F[iJ] = F
_

_
G[]e
i
R
d
d
x(x)J(x)

=0
. (2.40)
Expand this to get the perturbation expansion; we obtain the Feynman rules:
2.2.1 Feynman Rules for Interacting Scalar Field Theory
The Feynman rules to derive correlation functions between dierent points in spacetime are
A line between x and y represents a propagator i
F
(x y).
x y
A vertex with n lines represents a factor iV
(n)
(0), where V
(n)
() =
d
n
d
n
V ().

A
A
A
A

x
1
x
2
x
3
x
4
x
5
x
6
J(x) is represented by a vertex with one outgoing line.
, iJ(x)
Integrate over x for all vertices.
Introduce a symmetry factor S where necessary and divide by S.
If we consider an n-point correlation function
(x
1
) . . . (x
n
))
(n)
, (2.41)
we have diagrams with n external lines, one for each x
i
, and we drop J. The rst contributions
to (x)(y)) will be
x y +
x
`
_
y
z w
(S = 2)
There are slightly alternative versions of the Feynman rules:
Feynman Rules for Momentum Space
Consider
_
d
d
x
1
. . . d
d
x
n
e
i(p
1
x
1
+...+p
n
x
n
)
(x
1
) . . . (x
n
))
(n)
=: F
(n)
(p
1
, . . . , p
n
) , (2.42)
which will contain an overall delta function
d
(

i
p
i
) (energy-momentum conservation).
22

_
@
@R
@
@
p
2

p
1

p
3
@
@I
@
@
p
4
We make use of the Fourier transform of the Feynman propagator,
i
F
(x y) =
_
d
d
p
(2)
d
e
ip(xy)
_

i
p
2
+m
2
i
_
, (2.43)
and after integrating over the position for each vertex, generating a -function, this leads to
the following rules:
For each internal line with associated momentum k, add a propagator

i
k
2
+m
2
i
. (2.44)
-
k
For an external line with associated momentum p, add a propagator
i
p
2
+m
2
.
For each vertex with n outgoing lines, add a factor
iV
(n)
(0) (2)
d

d
_
i
p
i
_
. (2.45)
Integrate over momenta for all internal lines with measure
d
d
k
i
(2)
d
. (2.46)
The (2)
d
factor could have been associated with the propagator.
Thus for the loop diagram
-
p

_
-
p

-
-
k
1
k
2
the vertices will give factors
d
(p k
1
k
2
) and
d
(k
1
+k
2
p

), so after integration there will

be an overall delta function
d
(p p

).
Recall (1.133), Eulers formula for connected diagrams:
L = I V + 1 ,
where L is the number of loops, I is the number of internal lines and V is the number of vertices.
Let us look at the integrations performed for a particular diagram in momentum space. We
have the integral
_
I

i=1
d
d
k
i
(2)
d
V

v=1
(2)
d

d
_
i
p
i,v
_
, (2.47)
since there are I momenta associated to internal lines and V vertices where energy-momentum
conservation is imposed for all momenta p
i,v
for all lines incident at each vertex v. For
23
connected graphs there is one overall delta function for overall momentum conservation. After
removing all but one the momentum conservation -functions there are I V + 1 remaining
momenta to be integrated over, which is equal to L, the number of loops. Hence this becomes
(2)
d

d
_
i
p
i
_
L

j=1
d
d
k
j
(2)
d
, (2.48)
where the sum

i
runs over external lines and

j
is such that k
j
are L independent,
after imposing momentum conservation at each vertex, internal line momentum. The overall
(2)
d

d
_
i
p
i
_
is usually factored out.
Furthermore, if we look at the various factors of i in the Feynman integral, we notice that
there is a factor of i for each internal line and a factor of i for each vertex giving in total
(i)
I
(i)
V
= (1)
V
i
1L
. (2.49)
We can use this to simplify the Feynman rules in momentum space:
For each internal line with associated momentum k, add a propagator
1
k
2
+m
2
i
. (2.50)
For an external line with associated momentum p, add a propagator
i
p
2
+m
2
.
For each vertex with n outgoing lines, add a factor V
(n)
(0).
Impose momentum conservation at each vertex.
Add a factor i for each loop and integrate over undetermined loop momenta
_
d
d
k
(2)
d
.
There is an overall factor of i(2)
d
times a delta function imposing overall energy-
momentum conservation.
These rules may seem complicated, but to work them out in practice is quite straightforward.
Here are some illustrations.
For the simplest case n = 2, the diagram
-
p
1
p
2
corresponds to the amplitude
1
p
2
1
+m
2
(2)
d

d
(p
1
+p
2
) . (2.51)
For a potential with V
(4)
(0) = , the diagram
-
p
1
p
2

_ -
k
,
corresponds to
1
2
(i)
2
(p
2
1
+m
2
)(p
2
2
+m
2
)
i(2)
d

d
(p
1
+p
2
) ()(i)
_
d
d
k
(2)
d
1
k
2
+m
2
i
, (2.52)
where we have added a symmetry factor of 2.
For a potential V with V
(3)
(0) = g, the diagram
24
-
p
1
p
2

_ -

k
k p
1
is associated to the amplitude
1
2
(i)
2
(p
2
1
+m
2
)(p
2
2
+m
2
)
i(2)
d

d
(p
1
+p
2
) (g)
2
(i)
_
d
d
k
(2)
d
1
(k
2
+m
2
i)((k p
1
)
2
+m
2
i)
.
(2.53)
The guts of these calculations are in terms of doing the integrals. These are both one-loop
diagrams. When a diagram has no loops, there are no momentum integrations left.
In general, calculations with one or two loops are quite straightforward, but become rapidly
dicult for more loops.
2.2.2 Connected and Disconnected Graphs
A connected graph is one in which all lines are linked. Consider
(x
1
)(x
2
)(x
3
)(x
4
))
(4)
. (2.54)
In free theory, there are the following sets of graphs:
x
1
x
2
x
3
x
4
x
1
x
2
x
3
x
4

x
1
x
2
@
@
@
x
3
x
4
These are all disconnected graphs. In interacting theory, we have the following connected graphs:

x
1
x
2
@
@
@
x
3
x
4
,
@
@
x
1
x
2
x
3
x
4
Of course, there can still be disconnected graphs:
x
1
x
2

_
x
3
x
4
_
x
1
x
2
x
3
x
4
It is sort of obvious that any disconnected graph is composed of connected subgraphs. For
instance, in the above examples all disconnected graphs consist of two connected subgraphs.
In calculations, one can calculate disconnected graphs by calculating the connected subgraphs
and multiplying them together.
Let us now assume for simplicity that (x)) = 0, and write
(x
1
)(x
2
))
(2)
= (x
1
)(x
2
))
(2)
conn.
. (2.55)
By assumption, there are no graphs with one external line; we represent the sum of connected
graphs by

_
conn.
25
Now use this to decompose (x
1
)(x
2
)(x
3
)(x
4
))
(4)
into connected pieces. This will give
(x
1
)(x
2
)(x
3
)(x
4
))
(4)
= (x
1
)(x
2
))
(2)
(x
3
)(x
4
))
(2)
+ 2 similar terms
+(x
1
)(x
2
)(x
3
)(x
4
))
(4)
conn.
. (2.56)
We can represent this pictorially by
@
@

@
@
`
_
1
2 4
3
=
`
_
2
1
`
_
4
3
+
`
_
2
1
`
_
4
3
+

@
@
@
@
_
2
1
@
_
4
3
+
@
@

@
@
`
_
1
2 4
3
conn.
Now we want to discuss the generating functional which does this, i.e. which only gives con-
nected amplitudes. We previously dened Z[J] such that
(i)
n

J(x
1
)
. . .

J(x
n
)
Z[J]

J=0
= (x
1
) . . . (x
n
))
(n)
. (2.57)
If we now write Z[J] = e
iW[J]
, then W[J] is the generating functional for connected amplitudes.
We make the assertion that
(i)
n1

J(x
1
)
. . .

J(x
n
)
W[J]

J=0
= (x
1
) . . . (x
n
))
(n)
conn.
. (2.58)
The way to justify this is to show that if we substitute this formula into the rst expression, this
gives the correct expansion of an n-point correlation function . . .) in terms of connected m-
point correlation functions . . .)
conn.
which represent only connected diagrams. (All graphs with
n points can be expressed in terms of connected graphs with m n points.) Mathematically,
we express this relation as
(x
1
) . . . (x
n
))
(n)
=
n1

r=0

{i
1
,...,i
r
}{2,...,n}
(x
1
)(x
i
1
) . . . (x
i
r
))
(r+1)
conn.
(x
i
r+1
) . . . (x
i
n1
))
(nr1)
, (2.59)
where the second sum runs over all subsets of i
1
, . . . , i
r
2, . . . , n which contain r elements,
and we dene i
r+1
, . . . , i
n1
= 2, . . . , ni
1
, . . . , i
r
. This means that we pick one preferred
line corresponding to x
1
, and sum over all connected graphs including x
1
with between one and
n external lines, multiplied by all possible graphs for the remaining external lines. This gives a
recursive denition for the left-hand side.
In calculations like these, it is sometimes hard to see the wood for the trees; we will clarify
the mathematical expression by drawing appropriate pictures.

A
A
A
A

1
2
.
.
.
n
=
A
A
1
`
_

A
A
A
A
conn.
2
3
.
.
.
n
+
A
A
1
`
_
A
A
i

A
A
conn.
2
3
.
.
.
n
exclude i
26
+ . . . +

A
A
A
A

conn.
1
2
.
.
.
n
We will now show that assuming that W[J] generates all connected graphs, the relation Z[J] =
e
iW[J]
holds. We use the above relation, substituting in our expressions for the (r + 1)-point
correlation function and the connected (n r 1)-point correlation function in terms of Z[J]
and W[J]:

J(x
1
)
. . .

J(x
n
)
Z[J]

J=0
= i
n1

r=0

{i
1
,...,i
r
}{2,...,n}

J(x
1
)

J(x
i
1
)
. . .

J(x
i
r
)
W[J]

J=0

J(x
i
r+1
)
. . .

J(x
i
n1
)
Z[J]

J=0
=
i
(n 1)!

permutations
(2,...n)
n1

r=0
_
n 1
r
_

J(x
1
)

J(x
i
1
)
. . .

J(x
i
r
)
W[J]

J=0

J(x
i
r+1
)
. . .

J(x
i
n1
)
Z[J]

J=0
= i

J(x
2
)
. . .

J(x
n
)
__

J(x
1
)
W[J]
_
Z[J]
_

J=0
, (2.60)
where in the last line we have used the generalised Leibniz
20
rule,
d
n1
dx
n1
(f(x) g(x)) =
n1

r=0
_
n 1
r
_
d
r
dx
r
f(x)
d
nr1
dx
nr1
g(x) , (2.61)
extended to functional derivatives. After applying this the result is symmetric in x
2
, . . . , x
n
and the sum over permutations then just cancels the (n 1)! factor. Since the relation is true
for arbitrary n it shows that all terms in a Taylor
21
expansion of both sides of

J(x)
Z[J] = i
_

J(x)
W[J]
_
Z[J] . (2.62)
around J = 0 are the same and hence this holds for any J. Finally this is solved by
Z[J] = e
iW[J]
. (2.63)
This is the relationship between the generating functional for all n-point functions and the
corresponding functional for connected n-point functions.
The overall normalisation of Z[J] is generally chosen such that Z[0] = 1 (which basically
means that 1) = 1) and then W[0] = 0.
In momentum space, all connected amplitudes have an overall momentum conservation delta
function. You can always write
_
d
d
x
1
. . . d
d
x
n
e
i
P
i
p
i
x
i
(x
1
) . . . (x
n
))
(n)
conn.
= i(2)
d

d
_
i
p
i
_
(p
1
, . . . , p
n
) , (2.64)
where the function (p
1
, . . . , p
n
) is dened only for

i
p
i
= 0, and so basically is a function of
n 1 momenta. In the case n = 2, we have (p, p).
20
Leibniz, Gottfried Wilhelm von (1646-1716)
21
Taylor, Brook (1685-1731)
27
2.2.3 One Particle Irreducible Graphs
A connected graph is one particle reducible if it can be made disconnected by cutting one
(internal) line; otherwise it is one particle irreducible. For example, the graphs
A
A
A

A
A
A
A
A
A

A
A
A
are one particle reducible, whereas the graph
@
@

@
@
is one particle irreducible.
As we have seen it is convenient to consider connected graphs only. We notice that since all
one particle reducible graphs can be formed from one particle irreducible graphs, we can also
consider one particle irreducible graphs only.
A
A
A

A
A
A
=
A
A
A

A
A
A
Now construct a generating functional for one particle irreducible (short: 1PI) graphs; we
will call it . We seek to nd a relation between W and to have relations
Z[J] W[J] [] . (2.65)
The relationship between W and is a little more complicated than the one between Z and
W. Assume we have W[J], which is dened for any J(x); then a Taylor expansion will give
(x
1
) . . . (x
n
))
conn.
. We dene
(x))
J
=
W[J]
J(x)
=: (x) . (2.66)
We assume that there is an invertible relation between and J, so that we can express = (J)
and J = J() (this is non local so that (x) does not just depend on J(x)). Let us also assume
that if J = 0, then = 0. Then dene [] by
W[J] + [] =
_
d
d
x

(x

)J(x

). (2.67)
To see what this means, let us dierentiate with respect to (x):

(x)
W[J] +

(x)
[] = J(x) +
_
d
d
x

(x

)
J(x

)
(x)
. (2.68)
28
The rst term on the left is, by the standard chain rule,

(x)
W[J] =
_
d
d
x

W[J]
J(x

)
J(x

)
(x)
=
_
d
d
x

(x

)
J(x

)
(x)
. (2.69)
So it cancels with the second term on the right and we are left with

(x)
[] = J(x) . (2.70)
The relationship between W and is sometimes called a Legendre
22
transformation. (You
encounter this in thermodynamics, for example; consider the relationship F E between
free energy and energy. Given the energy E(S) as a function of entropy, we can dene the
temperature T by T =
E
S
and make a Legendre transformation
f(T) = E TS, (2.71)
so that
F
T
= S.)
[] is the generating functional of one particle irreducible graphs.
We need to nd a formula for

J(x)
in terms of

(y)
. We use the chain rule to obtain

J(x)
=
_
d
d
y
(y)
J(x)

(y)
=
_
d
d
y

2
W[J]
J(x)J(y)

(y)
= i
_
d
d
y G
2
(x, y)

(y)
, (2.72)
with the denition
G
n
(x
1
, . . . , x
n
) = (i)
n1

J(x
1
)
. . .

J(x
n
)
W[J] , (2.73)
so that G
n
(x
1
, . . . , x
n
)

J=0
= (x
1
) . . . (x
n
))
(n)
conn.
. We also introduce the notation

n
(x
1
, . . . , x
n
) = i

(x
1
)
. . .

(x
n
)
[] . (2.74)
Now we claim that [] generates one particle irreducible graphs, i.e.

n
(x
1
, . . . , x
n
)

=0
= (x
1
) . . . (x
n
))
conn.,1PI
. (2.75)
By using our previously derived formula, we note that

d
(x z) =
J(z)
J(x)
= i
_
d
d
y G
2
(x, y)

(y)
J(z)
= i
_
d
d
y G
2
(x, y)

2
(y)(z)
[] =
_
d
d
y G
2
(x, y)
2
(y, z) . (2.76)
Now, G
2
(x, y) and
2
(z, y) are essentially like matrices with continuous indices x, y, z; then

d
(x z) is essentially the unit matrix. This shows that
G
2
=
1
2
. (2.77)
For further calculations, introduce pictorial representations. Represent G
n
and
n
by

A
A
A
A

1
2
.
.
.
n
W/
22
Legendre, Adrien-Marie (1752-1833)
29
where we write W for G
n
and for
n
, respectively. We have shown that

_
=
_

_
_
1
W
Note also that from the above formula,

_
W
i

J(x)
=

(y)
x
y
Note that i

J(x)
adds an external line to G
n
, while

(y)
adds an external line to
n
. Therefore
i

J(x)

_
W =

_
W
=

_
W

(y)
_

_
_
1

_
W

_
W

_
W
where we have generalised the standard result for matrices
d
d
M
1
= M
1
d
d
M M
1
, (2.78)
to innite dimensions, so that

(w)
(
2
(y, z))
1
=
_
d
d
u d
d
v
2
(y, u)
1

2
(u, v)
(w)

2
(v, z)
1
=
_
d
d
u d
d
v
2
(y, u)
1

3
(w, u, v)
2
(v, z)
1
=
_
d
d
u d
d
v G
2
(y, u)
3
(w, u, v)G
2
(v, z) . (2.79)
This gives
G
3
(x, y, z) = i

J(x)
G
2
(y, z) =
_
d
d
u d
d
v d
d
w G
2
(x, w)G
2
(y, u)
3
(w, u, v)G
2
(v, z) , (2.80)
expressing G
3
in terms of
3
. This can be extended to higher values of n, for instance in
diagrammatic terms for n = 4,
i

J(x)

_
W =

_
@

@
W =
@
@

_
W

_
W

_
W
30
+ 2 others +

_
W

_
W

_
W

_
W
In the rst graph, and its two partners, we
can substitute for the W bubble with three
external lines using previous results. If this
is done all graphs on the right-hand side are
then one particle reducible graphs.
2.2.4 Tree Approximation
If Feynman graphs with loops are neglected there is a very simple expression for the generating
functionals W and in terms of the original action S. We start from the functional integral
result
Z[J] =
_
d[] e
i

(S[]+
R
d
d
x J(x)(x))
, (2.81)
where Plancks constant is reintroduced. This ensures the exponent is dimensionless if S has
the dimensions of action (ML
2
T
1
). In the classical limit when loops are discarded 0.
The functional integral is then dominated by contributions where the exponent is stationary or

(x)
S[]

=
J
= J(x) , (2.82)
and then
Z[J] e
i

(S[
J
]+
R
d
d
x J(x)
J
(x))
. (2.83)
Hence setting = 1 again in this limit for zero loops
W[J]
(0)
= S[
J
] +
_
d
d
x J(x)
J
(x) . (2.84)
It is easy to see that

J(x)
W[J]
(0)
=
J
, (2.85)
and therefore
[]
(0)
= S[] . (2.86)
2.3 Fermionic Fields
We proceed to fermionic
23
elds and will develop a formalism to dene a quantum eld theory
for fermions in terms of path integrals. The critical dierence to bosonic
24
elds is that
fermionic elds anti-commute, which is necessary for consistency with relativity and locality.
We know how to dene functional integrals for bosonic elds, and we want to construct a
framework where we can look at functional integrals in a similar fashion.
To do this we rst discuss functions of a nite set of anti-commuting or Grassmann
25
variables
i
, i = 1, . . . , n, which satisfy

j
+
j

i
= 0 , (2.87)
23
Fermi, Enrico (1901-1954), Nobel Prize 1938
24
Bose, Satyendra Nath (1894-1974)
25
Grassmann, Hermann G unther (1809-1877)
31
for all i and j. It follows that for all i

i
2
= 0 . (2.88)
Apart from anti-commuting Grassmann numbers form a vector space, they can be added and
multiplied by conventional numbers a so that a = a and since multipliction
i

j
is allowed
mathematically they form a ring.
It follows from
2
i
= 0 that any function f() is a nite linear sum
f() = a +a
i

i
+
1
2
a
ij

j
+. . . +
1
n!
a
i
1
...i
n

i
1

i
2
. . .
i
n
, (2.89)
where we can take the coecients to be totally antisymmetric, i.e. a
ij
= a
ji
, a
ijk
= a
jik
=
a
jki
etc.
We can also dene a dierentiation operator

i
that also anti-commutes

j
+
j

i
=
ij
. (2.90)
We further need to extend the notion of integration to Grassmann numbers, requiring that for
any function f() it is linear, translation invariant and gives an ordinary number depending on
f. (We only consider analogues to integrals over all x

dx f(x).) By translation invariance

we must have
_
d =
_
d ( +
0
) , (2.91)
which means that
_
d = 0 . (2.92)
We can choose a normalisation such that
_
d = 1 . (2.93)
In consequence, we have for any function
_
d (a +b) = b =

(a +b) , (2.94)
so dierentiation and integration is much the same, at least in this case. For n variables
i
we
dene an integration measure
d
n
= d
n
d
n1
. . . d
1
. (2.95)
Note that since
d
i
d
j
= d
j
d
i
, (2.96)
which is necessary for consistency, the order of the dierentials is important. So for a general
function f(),
_
d
n
f() = a
12...n
=
1
n!

i
1
i
2
...i
n
a
i
1
i
2
...i
n
, (2.97)
where
i
1
...i
n
is the n-dimensional antisymmetric symbol with
12...n
= 1.
26
Equivalently, we
can note that
_
d
n

i
1
. . .
i
n
=
i
1
i
2
...i
n
. (2.98)
It is easy to see that with these denitions
_
d
n

i
f() = 0 , (2.99)
and hence we can integrate by parts, taking into account the anti-commuting properties of the
derivative.
26
This is the higher-dimensional analogue of the famous
ijk
, of course.
32
For a change of variables

i
= A
ij

j
, where A is an (n n) matrix:
_
d
n
f(A) =
_
d
n
a
12...n
A
1i
1
. . . A
ni
n

i
1
. . . i
n
= a
12...n
A
1i
1
. . . A
ni
n

i
1
i
2
...i
n
= (det A) a
1...n
= det A
_
d
n
f() . (2.100)
Let us consider this in terms of

= A:
_
d
n
f(

) = det A
_
d
n

f(

) , (2.101)
so we obtain the transformation law
d
n

= d
n
(A) = (det A)
1
d
n
. (2.102)
Note that for bosonic variables,
d
n
(Ax) = det A d
n
x. (2.103)
det A is the Jacobian
27
which appears in this context.
2.3.1 Gaussian Integrals for Grassmann Variables
Consider integrals of the following form
_
d
n
e
1
2
A
ij

j
, (2.104)
where A is an antisymmetric (n n) matrix, i.e. A
ij
= A
ji
, where the dimension n is taken
to be even and we write n = 2m.
To evaluate this we expand the exponential; only the term containing n powers of give
a non zero contribution to the integral, i.e. we only need to consider the mth term in the
expansion:
_
d
n
e
1
2
A
ij

j
=
_
d
n

1
2
m
m!
A
i
1
i
2
. . . A
i
n1
i
n

i
1
. . .
i
n
=
1
2
m
m!
A
i
1
i
2
. . . A
i
n1
i
n

i
1
...i
n
Pf (A) , (2.105)
where Pf (A) is the Pfaffian
28
.
There is a relation between Pf (A) and det A: Consider a change of variables B =

and use the rules for a change of variables to obtain

Pf (A) =
_
d
n

e
1
2
A
ij

j
= (det B)
1
_
d
n
e
1
2
A
ij
B
ik

k
B
jl

l
= (det B)
1
Pf (B
T
AB) ,
(2.106)
(note that A
ij
B
ik
B
jl
= (B
T
AB)
kl
.) This implies
Pf (B
T
AB) = det B Pf (A) . (2.107)
A standard result for matrices is that we can nd a matrix B to put A in a standard form,
namely that (note that B
T
AB is antisymmetric for any matrix B):
B
T
AB =
_
_
_
_
_
_
_
_
_
_
_
0 1
1 0
0 1
1 0
.
.
.
0 1
1 0
_
_
_
_
_
_
_
_
_
_
_
=: J. (2.108)
27
Jacobi, Carl Gustav Jakob (1804-1851)
28
Pfa, Johann Friedrich (1765-1825)
33
Taking the determinant on both sides we get
(det B)
2
det A = 1. (2.109)
Furthermore,
Pf (J) =
1
2
m
m!
J
i
1
i
2
. . . J
i
n1
i
n

i
1
...i
n
=
1
2
m
m!
m!2
m
= 1 , (2.110)
therefore det B Pf (A) = 1. Eliminating det B from these relations we get
Pf (A)
2
= det A, Pf (A) =
_
det A. (2.111)
We may also extend this to the complex case: introducing Grassmann variables
i
and

i
,
i = 1, . . . , n, where

i
is the conjugate of
i
. We treat
i
and

i
as independent and assume the
rule
(
i
1
. . .
i
n
) =

i
n
. . .

i
1
. (2.112)
Furthermore all
i
anticommute with all

j
. Let B be a (n n) matrix and consider
_
d
n
d
n

i
B
ij

j
, (2.113)
where the integration measure is dened via
d
n
d
n

=
n

i=1
d
i
d

i
= d
1
d

1
. . . d
n
d

n
= d
n
d

n
. . . d
1
d

1
, (2.114)
for we can move pairs of Grassmann variables around without picking up minus signs. What
is nontrivial, slightly, is that

is placed to the right of here.
We expand the exponential and keep the nth term:
_
d
n
d
n

i
B
ij

j
=
_
d
n
d
n

1
n!
(

i
B
ij

j
)
n
=
1
n!

i
1
...i
n

j
1
...j
n
B
i
1
j
1
. . . B
i
n
j
n
= det B.
(2.115)
In general integration for Grassmann variables is more an exercise in algebra rather than
analysis and in some respects rather formal. However it plays a crucial part in the discussion of
quantum eld theories with fermion elds which include so-called ghost elds in gauge theories.
2.3.2 The Fermionic Oscillator
The harmonic oscillator for fermionic elds can be described by introducing creation and
annihilation operators

b

b which satisfy

b
2
= (

)
2
= 0,

b,

=

1. (2.116)
The states are built on a ground state [0) which satises

b[0) = 0 . (2.117)
We can dene
[1) =

b

[0) , (2.118)
so that

b

[1) = 0,

b[1) = [0). We have a two-dimensional space of states which can be expressed
in matrix form:
[0) =
_
1
0
_
, [1) =
_
0
1
_
;

b

=
_
0 0
1 0
_

b =
_
0 1
0 0
_
. (2.119)
We dene a Hamiltonian

H =

b =
_
0 0
0
_
. (2.120)
34
We now set up a path integral formalism to describe the system. It is convenient to introduce
a Grassmann variable . Dene states
[) = [0) +[1),

[ = 0[ +

1[ , (2.121)
and note that (
2
= 0 =

2
)

b[) = [),

[ . (2.122)
These are analogous to the states [z) which satisfy a[z) = z[z) for the bosonic oscillator. Now
we want to express the time evolution between these states

[ exp(i

Ht)[) = 0[0) +

1[e
it
[1) = 1 +

e
it
= exp(

e
it
) , (2.123)
in terms of a path integral by breaking up the time interval into small increments, just as we
did when we rst introduced path integrals. We need a completeness relation, which is given
by
_
d

d e

[)

[ =
_
d

d e

_
[0)0[ +

[1)1[
_
= [0)0[ +[1)1[ =

1, (2.124)
where we dropped mixed terms like

[0)1[ under the integral because they give no contribution.
We can use this to construct the path integral:

[ exp(i

Ht)[) =
_
d

[ exp(i

H(t t

))[

[ exp(i

Ht

)[) . (2.125)
This shows how to introduce an intermediate set of states to break up the time evolution into
smaller parts.
2.3.3 Path Integral Formalism for Fermionic Fields
We want to express the exact result for a time evolution amplitude

[ exp(i

HT)[
0
) = 1 +

0
e
iT
= e

0
e
iT
(2.126)
in terms of a path integral, proceeding in a fashion analogous to the bosonic case. We write
T = (N+1), being interested in the limit N , 0. We introduce N sets of intermediate
states and use the completeness relation (2.124) to obtain

[ exp(i

HT)[
0
) =
_
_
N

r=1
d

r
d
r
e

r
[ exp(i

H)[
r1
)
_

[ exp(i

H)[
N
)

_
N

r=1
d

r
d
r
e
iS
, (2.127)
where the exponential is
iS =
N

r=1
_

r
+

r1
e
i
_
+

N
e
i

r=1
_

r
+

r1
(1 i)
_
+

N
(1 i) , (2.128)
the approximation being valid for small . We can rewrite this as
S =
N

i=1
_
i

r1

r1
_
i

N
. (2.129)
Now we formally take the limit N , 0; the sum is essentially like the standard
denition of an integral:
S
T
_
0
dt (i

(t)

(t)

(t)(t)) i

(T) , (2.130)
35
making the replacements
r
(t
r
),

r


(t
r
), and where t
r
= r becomes a continuous
variable in the limit and we have the conditions (0) =
0
,

(T) =

. This is now written as
a path integral

[ exp(i

HT)[
0
) =
_
d[

] d[] e
iS[,

]
, (2.131)
where the integration runs over all fermion elds (t),

(t) with (0) =
0
,

(T) =

. This is
the fermionic path integral.
The path integral can be shown to give the required answer for this free example in a direct
way: expand ,

about a classical path for which the action S is stationary, treat and

as
independent elds. What are the equations that you get? If you vary

, you get
i

(t) = (t) , (2.132)
vary to get (using integration by parts)
i

(t) =

(t) . (2.133)
We can write down the classical solutions

c
(t) =
0
e
it
;

c
(t) =

e
i(tT)
. (2.134)
It is an exercise to substitute this in so that the result for the action is
S
c
S[
c
,

c
] = i

c
(T) = i

0
e
iT
. (2.135)
Now use the expansion
(t) =
c
(t) +(t),

(t) =

c
(t) +

(t) , (2.136)
and the fact that S is stationary for the classical path (s. above discussion for the bosonic
harmonic oscillator), to state that
S[,

] = S[
c
,

c
] +S[,

] . (2.137)
Then the path integral will become
_
d[

] d[] e
iS[,

]
= e
iS
c
_
d[

] d[] e
iS[,

]
. (2.138)
Note that the integral is independent of
0
,

and basically a constant so that it can be dened
to be one. This will yield the result
_
d[

] d[] e
iS[,

]
= e
iS
c
= 1 +

0
e
iT
, (2.139)
and we have reproduced the result derived before.
Note that we can rewrite
S[,

] =
T
_
0
dt

D i

(T) , (2.140)
where D = i
d
dt
+, which is essentially the Dirac
29
operator in one dimension. The integral
that we have dened to be one is then generally equal to det D.
Let us briey generalise this to eld theory; the essential idea was already contained in the
simple example before. We are interested in integrals of the following form
_
d[

] d[] e
iS
, S =
_
d
d
x

D , (2.141)
29
Dirac, Paul Adrien Maurice (1902-1984), Nobel Prize 1933, St. Johns
36
where the generalisation is that the Dirac operator in d dimensions is
D = +m1 , (2.142)
with the Dirac matrices

satisfying

= 2

. (2.143)
(Note that in the convention used herein,
00
= 1,
ij
=
ij
.) As before,
_
d[

] d[] e
iS
= det D, (2.144)
and for the free theory we may choose the normalisation such that det D = 1.
2.3.4 Canonical Treatment
For the simple fermionic oscillator the Lagrangian is just
L = i

, (2.145)
and we may dene the canonical momentum associated with by
p

=
L

= i

, (2.146)
and Hamiltonian
H = p

L =

. (2.147)
In a quantum treatment ,

become operators with anti-commutation relations at equal times
_

=

1. (2.148)
In this case

H =

, (2.149)
and
_

H,

=

,
_

H,

. (2.150)
2.3.5 Propagators
This is the situation for free elds. We dene a propagator
iS
F
(x y)

:=

(x)

(y)) =
1
det D
_
d[

] d[]

(x)

(y) e
iS
, (2.151)
where S
F
is the Feynman propagator for fermions. Let us follow the same procedure which
we considered in the bosonic case to derive the form of the propagator. Apply the Dirac
operator to the propagator, suppressing spinor indices:
D
x
(x)

(y)) =
1
det D
_
d[

] d[]
S

(x)
e
iS

(y)
=
1
det D
_
d[

] d[] (i)
_

(x)
e
iS
_

(y)
=
1
det D
_
d[

] d[] ie
iS

(x)

(y)
= i 1
d
(x y) , (2.152)
where we have used the fact that D
x
(x) =

(x)
S and integrated by parts. We need to solve
the dierential equation
D
x
S
F
(x y) = 1
d
(x y) , (2.153)
37
and therefore use Fourier transforms, as before. The Fourier transform is dened by

S
F
(p) =
_
d
d
x e
ipx
S
F
(x) , (2.154)
which gives the equation for

S
F
(i p m1)

S
F
(p) = 1 . (2.155)
To solve that we make use of the identity
(i p +m1)(i p m1) = ( p)
2
m
2
1 = (p
2
m
2
)1 , (2.156)
so that
(p
2
m
2
)

S
F
(p) = i p +m1 , (2.157)

S
F
(p) =
i p +m1
p
2
m
2
+i
. (2.158)
The diagrams representing propagators are drawn as before, with a propagator

(x)

(y)) = iS
F
(x y)

(2.159)
being represented by a line

x y

while a propagator

(x)

(y)) = iS
F
(y x)

(2.160)
is represented by
-
x y

Besides the Fermion elds by themselves it is necessary to consider operators formed by
products such as

M where M is some Dirac matrix. The electromagnetic current for charged
Fermions is of this form with M

. We may then consider correlation functions involving

such operators at dierent points. Even for free elds these are more complicated. For the
simplest case of two such operators we obtain

(x)M(x)

(y)M

(y)) = tr(MS
F
(x y)M

S
F
(y x)) (2.161)
using the above results for the propagators for (x)

(y) and (y)

(x) where it is also nec-

essary to include also a minus sign to take account of the anti-commuting properties of the
Fermion elds. In calculating this result we neglect contributions involving the propagator for
(x)

(x), proportional to S
F
(0) which is divergent. In terms of Feynman diagrams this result
is represented by

_
, ,
-

y xM
which describes a Fermion loop. In general for any Feynman diagram with closed Fermion loops
the Feynman rules require that there is an additional minus sign for each such loop.
38
3 Operator Aspects of Quantum Field Theory
Many aspects of quantum eld theories require and understanding of how the elds and opera-
tors constructed from them act on the space of states formed by all multi-particle states. It is
also necessary to understand the role of symmetries.
3.1 Symmetries and Noethers theorem
Continuous symmetries, which form Lie groups, lead to relations between correlation functions
of elds. To derive these we rst consider a crucial result in classical dynamics. Noether
30
showed that when an action has a continuous symmetry, there is a conserved current and a
corresponding conserved charge.
Proof
Assume that there is a continuous symmetry of the classical theory. The symmetry transfor-
mations act on the elds such that they transform as

= +

, (3.1)
for an innitesimal parameter. If this is a symmetry this means that the action is invariant

S[] = 0 . (3.2)
Now allow (x) to be a function of x; in that case

S[] =
_
d
d
x

(x)j

(x) (3.3)
for some j

, any contribution must involve a derivative of since for a constant the action
is invariant (we can always discard total derivatives under the integral, this may introduce an
ambiguity in j

but this is not crucial.) If the equations of motion are obeyed then

S[] = 0 , (3.4)
for any because this the action is stationary for arbitrary when satises its equations of
motion. Hence we must have

= 0 (3.5)
subject to the equations of motion. A conserved charge is then given by
q =
_
d
d1
x j
0
(x), (3.6)
since

Q =
_
d
d1
x

j(x) = 0.
In a quantum treatment the charge becomes an operator such that
_

Q,

(x)

= i

(x) . (3.7)
As an illustration let L =

V (

). This has a symmetry corresponding to U(1)

transformations

= i,

= i

and the corresponding conserved current is then

j

= i(

). In this case
_

Q,

(x)

(x).
3.2 Ward Identities
In a quantum eld theory there are Ward
31
identities associated with symmetries for correlation
functions. In general we consider
X) =
1
Z
_
d[] X()e
iS[]
, (3.8)
30
Noether, Emmy (1882-1935)
31
Ward, John Clive (1924-2000), unfortunately no Nobel Prize
39
where X() is a function of the elds, e.g. X() = (x
1
)(x
2
) . . .. Let us suppose that under
the transformation

,
X X

= X(

) = X +

X , (3.9)
and since the functional integral does not depend on the choice of variable
_
d[

] X

e
iS[

]
=
_
d[] X e
iS[]
. (3.10)
We then expand the left-hand side in , assuming invariance of the measure d[

] = d[], with
the implications,
_
d[

] X

e
iS[

]
=
_
d[] (X +

X)e
iS[]
_
1 i
_
d
d
x

_
+O(
2
) , (3.11)
and this then requires

X) = i
_
d
d
x

(x)j

(x) X) . (3.12)
Taking a functional derivative with respect to (x) on both sides, we obtain a Ward identity

(x)

X) = i

(x) X) . (3.13)
Ward rst emphasised this in a particular example in quantum electrodynamics where it played
a crucial role in understanding renormalisation since it showed two divergent renormalisation
constants were equal. Integrating over x, we obtain
_
d
d
x

(x)

X) = 0 , (3.14)
or for constant ,

X) = 0. (3.15)
This argument, however, is by no means watertight. It may fail for two reasons, which are
essentially related;
(i) we assumed invariance of the measure d[] = d[

];
(ii) the functional integral is not dened without regularisation, and you have to check whether
the regularisation also satises the symmetries otherwise this might lead to anomalies.
3.3 Energy Momentum Tensor
In any relativistic quantum eld theory the energy momentum tensor plays a crucial role. From
it the momentum operator, the angular momentum operators and the generators of Lorentz
boosts can be constructed. The Noether procedure can be adapted to determine the form of the
energy momentum starting from an action which is invariant under translations and Lorentz
transformations.
First we consider just translations. For a translation invariant theory the action is invariant
under under innitesimal translations so that

a
= a

a
S[] = 0 , (3.16)
for constant innitesimal d-vectors a

. If a

is allowed to depend on x this implies

a
S[] =
_
d
d
x

(x) T

(x) , (3.17)
and as before when the equations of motion are satised so that S is stationary then T

is
conserved and the momentum operator is dened by a spatial inteegral

= 0 , P

=
_
d
d1
x T
0

(x) . (3.18)
40
However in relativistic quantum eld theories the energy momentum tensor should be sym-
metric

, as required for coupling to gravity. To construct

from S we must also

consider Lorentz transformations. In general these act on the elds according to

+
1
2

, (3.19)
where s

= s

are the spin generators acting on . These are matrices satisfying the
commutation relations
[s

, s

] =

, (3.20)
so that
(

) =
[

,]
, [

, ]

. (3.21)
Lorentz and translation invariance ensures that the lagrangian density L is a scalar so that

v
L(, ) = v

L(, ) , v

(x) = a

, (3.22)
where
v
is dened by combining
a
and

and
v
= (
v
). Note that

(x) +

(x) = 0 , (3.23)
so that v

(x) is a Killing vector. For general v

(x) and

(x) =

(x) then we can write

v,
S[] =
_
d
d
x
_
(

) T

_
, X

= X

, (3.24)
so that when v

= a

with a

constants then S is invariant. This determines T

but it is not in general symmetric. However if we dene

= T

(X

+X

) , (3.25)
then since
_
d
d
x

(X

+X

) =
_
d
d
x (

)X

= 0
_
d
d
x

(X

+X

) =
_
d
d
x

, (3.26)
the invariance condition becomes

v,
S[] =
_
d
d
x (

. (3.27)
For satisfying the equations of motion S[] = 0 for any so that then
v,
S[] = 0. Since
v

(x),

(x) are arbitrary this gives, subject to the equations of motion,

= 0 ,

. (3.28)
If we dene
/

= x

, (3.29)
then the symmetry and conservation equations imply

= 0 . (3.30)
Alternatively let
J

v
= v

v
= 0 for v

a Killing vector . (3.31)

This allows us to dene the Lorentz generators by
M

=
_
d
d1
x /
0
. (3.32)
41
Ward identities for the energy momentum tensor can be derived in a similar fashion as for
conserved currents. It is convenient rst to assume the energy momentum tensor is symmetric
and require

v
S[] =
_
d
d
x

for
v
= v

1
2

. (3.33)
Then the Ward identities come from

v
X) = i
_
d
d
x

(x)

(x) X) . (3.34)
If v

is a Killing vector this just says that

v
X) = 0. For general v

(x) we may obtain the

Ward identity

(x)

v
X) = i

(x) X) . (3.35)
3.4 Relation of n-Point Functions and Scattering Amplitudes
We go back to the case of just bosonic elds, having dealt with fermionic elds in the last
section. Consider the identity
(x
1
) . . . (x
n
)) = 0[T

(x
1
) . . .

(x
n
)[0) , (3.36)
where

denotes the Heisenberg
32
operator elds.
We would like to be able to talk about scattering amplitudes, which are measurable quantities:
2 particles many. (3.37)
First of all, we need to nd a representation for the case of two particles (x
1
)(x
2
)), which
incorporates standard eld theory and Lorentz invariance. We use operator methods and
look at the operator elds

. We have the momentum operator

= (

H,

P) , (3.38)
with the commutator
[

,

(x)] = i

(x) . (3.39)
In a sense, you can always solve this in terms of the eld at the origin

(x) = exp(i

P x)

(0) exp(i

P x) . (3.40)
We assume to have a complete set of states, such that

1 =

n
[n)n[ . (3.41)
Consider the following quantity (for a given d-vector p):
_
d
d
x 0[

(0)

(x)[0)e
ipx
=

n
_
d
d
x 0[

(0)[n)n[

(x)[0)e
ipx
=

n
_
d
d
x e
i(pP
n
)x
0[

(0)[n)n[

(0)[0)
= (2)
d

d
(p P
n
)[n[

(0)[0)[
2
, (3.42)
where

P

[0) = 0 and

P

[n) = P

n
[n) are used. The summation is constrained by virtue of
the delta function. There will only be a non-zero contribution if p
0
> 0 since the states have
positive energy; furthermore it is a Lorentz scalar and it is convenient to write this as
_
d
d
x 0[

(0)

(x)[0)e
ipx
= 2(p
0
)(p
2
) , (3.43)
32
Heisenberg, Werner (1901-1976), Nobel Prize 1932
42
where (p
0
) is a step function and (p
2
) is a function which can only be non-zero for p
2
> 0,
since P
n
2
< 0. We also have (p
2
) 0.
In a corresponding fashion we also have
_
d
d
x 0[

(x)

(0)[0)e
ipx
= (2)
d

d
(p +P
n
)[n[

(0)[0)[
2
= 2(p
0
)(p
2
) . (3.44)
These results then allow us to obtain a formula for the time-ordered product by inverting the
Fourier transform:
0[T

(0)

(x)[0) = 2
_
d
d
p
(2)
d
e
ipx
(p
2
)
_
(x
0
)(p
0
) +(x
0
)(p
0
)
_
, (3.45)
which combines the two separate cases for x
0
> 0 and x
0
< 0 together. In order to separate
the dependence on we introduce another integration variable so that:
0[T

(0)

(x)[0) = 2

_
0
d ()
_
d
d
p
(2)
d
e
ipx
(p
2
+)
_
(x
0
)(p
0
) +(x
0
)(p
0
)
_
.
(3.46)
It is now possible to carry out the p
0
integration using the delta function; we get a contributions
for p
0
= E
p
:=
_
+ p
2
, but only one contributes in each term because of the step function.
This gives
0[T

(0)

(x)[0) =

_
0
d ()
_
d
d1
p
(2)
d1
e
i px
2E
p
_
(x
0
)e
iE
p
x
0
+(x
0
)e
iE
p
x
0
_
. (3.47)
Note that this expression is almost identical to our expression for the Feynman propagator for
free scalar eld theory, so reiterating the calculation we did then we get
0[T

(0)

(x)[0) =

_
0
d () (i)
_
d
d
p
(2)
d
e
ipx
p
2
+ i
. (3.48)
This is a sum of Feynman propagators for dierent values of m
2
.
The formula is valid for any eld theory incorporating the assumption of Lorentz invariance
and so on.
We now determine the contribution of a single particle to the sum over states. Remember
that the function (p
2
) was dened by
2(p
0
)(p
2
) =

n
(2)
d

d
(p P
n
)[n[

(0)[0)[
2
. (3.49)
Let [

P) be a single-particle state, satisfying P

2
= m
2
and P
0
> 0. We use the convention for
normalisation of single-particle states

P) = (2)
d1
(2E

P
)
d1
(

P) . (3.50)
Suppose that

P[

(0)[0) = N, (3.51)
where N is a scalar that only depends on P
2
= m
2
, and so is essentially independent on
P. The contribution to the sum which results from this particular particle may be isolated
explicitly:

n
[n)n[ =
_
d
d1
P
(2)
d1
1
2E

P
[

P)

P[ +. . . . (3.52)
43
To nd the contribution to (p
2
), consider the following integral, obtained by inserting the
above relation into the dening relation for :

n
(2)
d

d
(p P
n
)[n[

(0)[0)[
2
=
_
d
d1
P
(2)
d1
1
2E

P
(2)
d

d
(p P)[

P[

(0)[0)[
2
= N
2
(2)
1
2E
p
(p
0
E
p
)
= N
2
2(p
2
+m
2
)(p
0
) , (3.53)
where E
p
=
_
m
2
+ p
2
. This requires (p
2
) = (p
2
+m
2
)N
2
or
() = ( m
2
)N
2
. (3.54)
What implication does this have for the time-ordered expression? Consider
_
d
d
x 0[T

(0)

(x)[0) e
ipx
= i

_
0
d()
1
p
2
+ i
i
N
2
p
2
+m
2
+. . . , (3.55)
where . . . includes contributions whose singularities arise for p
2
> m
2
. In general n-particle
states contribute to () only when (nm)
2
and these give rise to branch cuts in (3.55) with
branch points at p
2
= (nm)
2
.
Hence single particle states give rise an isolated pole at p
2
= m
2
, which we express as this as

_

P P
i
N
2
p
2
+m
2
,
where means that only single-particle contributions are considered.
Let us show how we can now relate the amplitudes including elds to scattering amplitudes.
Consider the integral
_
d
d
x[

(x)[)
conn.
e
ipx
, (3.56)
corresponding to the picture

_

_
6
p
Fourier transform
of (0)(x))
We have an external line corresponding to

, with an associated momentum p. The amplitude
contains a factor
_
d
d
x (0)(x)) e
ipx
which corresponding to all one particle reducible graphs
connected to this external line. As we have seen, this has a pole for p
2
m
2
resulting from
the contribution of the single-particle state [ p), which is only dened for physical momenta such
that p
2
= m
2
, to the sum over states. Acting on the vacuum state, to the right or to the left,
the eld

, in a time ordered product, thus creates a single particle state so that
_
d
d
x

(x)[0) e
ipx

i
p
2
+m
2
[ p) p[

(0)[0) , p
0
> 0 ,
_
d
d
x0[

(x) e
ipx

i
p
2
+m
2
0[

(0)[ p) p[ , p
0
< 0 , (3.57)
44
as p
2
m
2
. A similar result holds also when

acts on any state [) to the right, giving
[, p), or equivalently on [ to the left giving then , p[. Hence for the full amplitude there
are poles at p
2
= m
2
corresponding to an additional incoming or outgoing particle, with
momentum p or p, so that
_
d
d
x [

(x)[)
conn.
e
ipx

_
i[, p)
1
p
2
+m
2
p[

(0)[0) = i[, p)
N
p
2
+m
2
if p
0
> 0 ,
i0[

(0)[ p)
1
p
2
+m
2
, p[) = i, p[)
N
p
2
+m
2
if p
0
< 0 ,
as p
2
m
2
.
(3.58)
Alternatively this can be written as
[, p) = lim
p
2
m
2
i(p
2
+m
2
)
N
_
d
d
x [

(x)[) e
ipx
, p
0
> 0 , (3.59)
with a similar result for , p[) when p
0
< 0. This can be extended to arbitrarily many
initial or nal particles. For four the situation looks like this:
@R
@
_
@
@
p
1

p
2
@I
@
_
@
@
p
4

p
3

_
The analogous expression is
_
4

i=1
d
d
x
i
e
ip
i
x
i
(x
1
)(x
2
)(x
3
)(x
4
))
conn.
= i(2)
d

d
(

i
p
i
)
4
(p
1
, p
2
, p
3
, p
4
) . (3.60)
We claim that the scattering amplitude for physical particles is given by (e.g.for p
0
1
, p
0
2
> 0 and
p
0
3
, p
0
4
< 0)
p
3
, p
4
[

T[ p
1
, p
2
) =
i
4
N
4
lim
p
2
1
,p
2
2
,p
2
3
,p
2
4
m
2
4

i=1
(p
2
i
+m
2
)
4
(p
1
, p
2
, p
3
, p
4
) . (3.61)
45
4 Calculation of Feynman Amplitudes and Divergences
This is a very broad subject, and there are very complicated methods for very complicated
diagrams. We can only touch the surface here. Let us start with a Lagrangian density
L =
1
2

V (). (4.1)
The Feynman rules for this theory in momentum space were derived above:
For each internal line with associated momentum k, add a propagator
1
k
2
+m
2
i
. (4.2)
For an external line with associated momentum p, add a propagator
i
p
2
+m
2
.
For each vertex with n outgoing lines, add a factor V
(n)
(0).
Impose momentum conservation at each vertex.
Add a factor i for each loop and integrate over undetermined loop momenta
_
d
d
k
(2)
d
.
There is an overall factor of i(2)
d
times a delta function imposing overall energy-
momentum conservation.
For any graph, we denote the number of internal and external lines by I and E, the number
of n-vertices by V
n
, the total number of vertices by V , so that V =

n
V
n
, and the number of
loops by L.
4.1 Two Key Ideas
First of all let us describe a useful trick which simplies things, the so-called Wick
33
rotation.
4.1.1 Wick Rotation
Note that all propagators have singularities at k
2
+ m
2
= 0. The actual Feynman integrals
dene analytic functions, and one should think of all variables they depend on as being poten-
tially complex. When we get poles in the propagator, these correspond to single-particle states;
branch cuts correspond to multi-particle states.
Let us analyse the nature of these poles:
1
k
2
+m
2
i
=
1
(k
0
)
2
+ (E i)
2
, E =

k
2
+m
2
, (4.3)
where has been slightly redened. This expression has poles at k
0
= (E i), as we can
display in the complex k
0
plane:
- - -
6
6
6
,
,
new contour
33
Wick, Gian-Carlo (1909-1992)
46
We can rotate the k
0
contour in a standard way to the imaginary axis
K
0
ik
d
, where k
d
is real . (4.4)
You do this for all loop momenta; as long as there are no contributions at innity you get the
same answer. The propagator becomes
1
k
2
0
+

k
2
+m
2

1
k
2
d
+

k
2
+m
2
,
d
d
l
i(2)
d

d
d1
l dl
d
(2)
d
, (4.5)
where l is any loop momentum. We can apply this to any integration and it simplies calcula-
tions since all singularities have been removed. However there are slight caveats.
For the rotation to be consistent, we also have to rotate the external momenta to ensure
momentum conservation, so we must require
(p
E
i
)
2
> 0 , (4.6)
for each external line, and more generally for any subset of external lines
(

p
E
i
)
2
> 0 . (4.7)
Wick rotation changes Feynman amplitudes to integrals over Euclidean
34
space. We avoid
discussing singularities when we want to determine the form of some amplitude.
Wick rotations go from Minkowski to Euclidean space.
4.1.2 Appearance of Divergences in Feynman Integrals
Divergences may arise in integrating over L undetermined loop momenta according to the
Feynman rules. The basic integral we have to consider is (very schematically)
F =
_
d
dL
l
(2)
dL
1
k
2I
, (4.8)
where k = k(l) is typically linear in l, i.e. k l + other momenta. When does such an integral
diverge? Consider for > 0 the nite integral

_
0
dx
(x +)
+1
=
1

. (4.9)
This integral is divergent for 0 and the right-hand side expression has a pole at = 0. We
dene the degree of divergence of an integral F by
D = dL 2I . (4.10)
Evidently, if D 0, then the integral F is divergent. In order to rewrite D in a slightly dierent
form, we use Eulers formula (1.133), that is
L = I V + 1 ,
together with the condition

n
nV
n
= 2I +E , (4.11)
which states that each internal line is connected to two vertices and each external line is con-
nected to one vertex. First we eliminate I from the denition of D in (4.10) to obtain
D = dL

n
nV
n
+E , (4.12)
34
Euclid of Alexandria (c. 330 BC-c. 275 BC)
47
and then combining (4.11) and (1.133) we nd
2L =

n
nV
n
2V E + 2 . (4.13)
Now assuming d 4, we add (4.12) and twice (4.13) to get the nal result
D = (d 4)L +

n
(n 4)V
n
E + 4 . (4.14)
If the number of dimensions d is indeed equal to four, and we only have an interaction term
V () =
1
24

4
, then
D = 4 E , (4.15)
so D 0 for 0 E 4; graphs with more than four external lines are supercially nite,
there may however still be divergences from subgraphs. The important thing to notice here is
that D is independent of the number of loops; this will result in a renormalisable theory. (The
meaning of this word will become apparent in due course.)
In the situation d 3, we get the slightly dierent formula
D = (d 3)L +

n
_
1
2
n 3
_
V
n

1
2
E + 3 . (4.16)
You can combine the formulae in various ways at your personal convenience, and here we have
combined them in a way that brings out the point we want to make. For d = 3 and only a
6
interaction we get
D = 3
1
2
E , (4.17)
that is a divergence for not more than six external lines. Just as in the previous case, the theory
will be renormalisable.
We can extend this to more complicated theories, e.g.
L
2
+
4
, (4.18)
which contains vertices
@
@

k
@
@
@
@

where a vertex of the rst type has an associated momentum k because of the derivative
appearing in the Lagrangian. (We need not worry about the details for this point of view.)
The basic integral is then of the following form:
_
d
dL
l
(2)
dL
k
V
3
k
2I
. (4.19)
The denition for the quantity D is now modied from (4.10) to become
D = dL +V
3
2I , (4.20)
since each three-edged vertex contributes a power of the momentum to the integral F; elimi-
nating I, we nd the analogy of (4.12) is
D = dL 2V
3
4V
4
+E . (4.21)
The second equation we need is (4.13), which in our special case now reads
2L = V
3
+ 2V
4
. .
P
n
nV
n
2V
E + 2 . (4.22)
48
We can combine these together to obtain
D = (d 4)L + 4 E , (4.23)
and so D = 4 E in the case d = 4. This corresponds to a gauge theory.
For completeness, let us give one example of a non-renormalisable theory: Consider an
interaction of the following form
L
n2
()
2
, n = 2, 3 . . . , (4.24)
which is indeed what you get when you apply the standard rules of quantisation to gravity. We
have vertices with an arbitrary number of lines, with two associated momenta:
@
@
@
@

kk
The basic integral
_
d
dL
l
(2)
dL
k
2V
n
k
2I
(4.25)
has the degree of divergence
D = dL + 2V
n
2I = (d 2)L + 2 . (4.26)
Any graph in this example is divergent if d 2, and D increases with L.
A few comments:
For a renormalisable theory, if vertices of a certain type are present, e.g. V
3
for d = 4, the
degree of divergence is reduced. But for other types of vertices, e.g. if V
5
,= 0, the theory
becomes non-renormalisable.
It is possible that D < 0 for a graph but that there are subgraphs with D 0; then there
are sub-divergences. For instance, consider
4
theory in four dimensions and a graph with
external lines, so that D = 2:
H
H
H

H
H
H

_
There are subgraphs
@
@
@

_
@
@
@

which have D = 0 and are divergent.

We now consider
4
theory in four dimensions and one- and two-loop graphs. The number
of external lines E must be even for any graph in this theory. Vacuum graphs (E = 0) may be
disregarded. Hence it is sucient to consider just graphs with two and four external lines to
get all divergences.
The one-loop graphs we nd are
49
@
@
@

_
@
@
@

_
whilst the two-loop graphs with E = 2 are

_
`
_ `
_
`
_
`
_
where the last one is one-particle reducible corresponding to two one-loop integrals; and nally
for E = 4 we get
@
@
@

_
@
@
@

@
@
@

_
@
@
@

`
_
4.2 Regularisation
The critical thing to realise is that divergent integrals are meaningless; for unambiguous cal-
culations, it is necessary to remove the divergences. They arise when the momentum goes to
innity, so we need to suppress the large k contribution to the integral. There is no unique
prescription how to do this, and in the end, the exact method does not matter since they all lead
to the same result. Dierent methods however may be more or less useful in doing calculations.
A simple procedure is introducing a momentum cut-o, after Wick rotation,
K
2
<
2
(4.27)
in all integrals. This is not very good in general because it produces more complicated integrals.
Furthermore it is not systematic and does not respect the translation invariance (k k + k
0
)
of innite range integrals.
There are some desirable requirements for a regularisation scheme. A good regularisation
should
(i) be valid for arbitrary Feynman graphs, to all orders of perturbation theory;
(ii) respect the basic symmetries of the theory (this is very desirable), e.g. Lorentz invari-
ance;
(iii) ideally be applicable outside perturbation theory since quantum eld theories are not just
a perturbation expansion.
We will now consider one particular process which will work at least in some theories. We
modify the Lagrangian
L =
1
2

P
_

2
_

V () , (4.28)
where P is some polynomial satisfying P(0) = 1. Suppose that the potential is of the form
V () =
1
2
m
2

2
+O(
3
) . (4.29)
50
Note that the propagator is the inverse of the quadratic part of the Lagrangian, so we can
actually write down what the propagator is. After a Fourier transform which transforms

2
k
2
, we get
1
k
2
P(
k
2

2
) +m
2

1
(k
2
)
h+1
for large k , (4.30)
where we suppose that P is a polynomial of degree h. As , we recover the original
propagator
1
k
2
+m
2
. (4.31)
For a nite value of the divergences of Feynman integrals are suppressed. To demonstrate
this, we can consider a simple example; we set m = 0 and P(
k
2

2
) = 1 +
k
2

2
and obtain the
propagator
1
k
2
_
1 +
k
2

2
_ =
1
k
2

1

2
1
1 +
k
2

2
. (4.32)
The rst part of that corresponds to a (physical) particle; the second part has a pole at k
2
=
2
with negative residue, and so it is not a physical particle! Because of the wrong sign appearing
in the propagator, you will get negative norm states and thus the theory with P ,= 1 is not a
physical theory. We need to take the limit to get a physical theory.
This method is complicated to calculate.
4.2.1 Dimensional Regularisation for One-Loop Graphs
Another method that is frequently used is dimensional regularisation: Consider the theory in
d dimensions, where the number d is a complex parameter. We can extend Feynman graphs
to arbitrary d, and the critical thing is that we will nd that D < 0 for suciently small d. We
will dene the Feynman integrals by analytic continuation in d.
Let us rst consider a more mundane example which will be useful for later calculations.
The Gamma function is dened by
() =

_
0
dx x
1
e
x
. (4.33)
This is nite (i.e. well-dened) if Re > 0 and diverges for Re 0. But there is a standard
procedure by which we can extend the denition to allow for an analytic continuation: rewrite
the integral as
() =
1

_
0
dx
_
d
dx
x

_
e
x
=
1

_
x

e
x

0
+
1

_
0
dx x

e
x
=
1

( + 1) , (4.34)
where we used integration by parts and assumed the surface term vanished since Re > 0. But
since ( + 1) is given by a nite integral for Re > 1, we can use this relation to extend
the denition of () to the region Re > 1. Since (1) = 1, we get
()
0
1

. (4.35)
There is clearly a pole at = 0 which reects the divergence in the original integral. For
applications later it is worth noting that the Gamma function may be expanded
ln( + 1) =

k2
(1)
k
1
k
(k)
k
, (4.36)
where is Eulers number and (k) =

n1
n
k
.
We now calculate the Feynman integrals for specic one and two loop Feynman graphs
focusing on those where there are divergencies. We consider only connected one-particle irre-
ducible graphs, which are related to the generating functional []. We seek to calculate
n
as
51
dened by,
_
n

i=1
d
d
x
i
e
ip
i
x
i
(x
1
) . . . (x
n
))
conn.,1PI
= i(2)
d

d
(

i
p
i
)
n
(p
1
, . . . , p
n
) . (4.37)
The Feynman rules for the
n
are
Consider all connected one-particle irreducible graphs with n external lines.
For each internal line there is a factor
1
k
2
+m
2
.
For each n-vertex add a factor V
n
(0) and impose conservation of momenta for all lines
meeting at the vertex.
Integrate over the independent loop momenta with
(i)
_
d
d
l
(2)
d

_
d
d
l
(2)
d
. (4.38)
after Wick rotation.
Add the appropriate symmetry factor
1
S
for the graph.
Note that there are no additional factors for external lines. For divergent Feynman integrals
we need only consider n 4 and the calculations are manageable just for one and two loops.
Suppose we have a Lagrangian of the form
L =
1
2

1
2
m
2

1
6
g
3

1
24

4
. (4.39)
(The prefactors appearing in front of the interaction terms correspond to dividing by the order of
the permutation group for each term; they ensure that the symmetry factors for each Feynman
diagram are calculated as described previously.) For zero loops the non zero results are just
(using superscripts to denote the number of loops)

(0)
2
(p, p) = p
2
m
2
, (4.40a)

(0)
4
(p
1
, p
2
, p
3
, p
4
) = , (4.40b)

(0)
3
(p
1
, p
2
, p
3
) = g . (4.40c)
Consider rst one-loop diagrams for n = 2; one contribution arises from the diagram
-
p p

_ -
k
,

According to the Feynman rules, it will be given by

(1)
2
(p, p) =

2
_
d
d
k
(2)
d
1
k
2
+m
2
. (4.41)
This is a Euclidean integration and the integrand has no angular dependence, which means
we can replace the integration measure by
d
d
k S
d
[k[
d1
d[k[ , (4.42)
where S
d
is a constant prefactor given by
S
1
= 2 , S
2
= 2 , S
3
= 4 , S
4
= 2
2
, . . . , S
d
=
2
1
2
d
(
d
2
)
. (4.43)
52
To justify the result for general d consider the integral
_

_d
2
=
_
d
d
k e
k
2
= S
d

_
0
dk k
d1
e
k
2
k
2
=u
=
1
2
S
d

_
0
du u
d
2
1
e
u
= S
d
(
d
2
)
2
d
2
, (4.44)
set = 1 to obtain the result. In consequence, we can say that

(1)
2
(p, p) =

(
d
2
)
1
(4)
d
2

_
0
dk k
d1
1
k
2
+m
2
. (4.45)
Now in order to work out this integral, we need another trick; we can rewrite
1
k
2
+m
2
=

_
0
d e
(k
2
+m
2
)
. (4.46)
The integrals in this example are then just those dening the Gamma function:

(1)
2
(p, p) =

(
d
2
)
1
(4)
d
2

_
0
d

_
0
dk k
d1
e
(k
2
+m
2
)
k
2
=u
=

2(
d
2
)
1
(4)
d
2

_
0
d e
m
2

_
0
du u
d
2
1
e
u
=

2
1
(4)
d
2

_
0
d e
m
2 1

d
2
. (4.47)
Hence we obtain the nal result for this one loop graph

(1)
2
(p, p) =
1
2

(4)
d
2

_
1
d
2
_
(m
2
)
d
2
1
. (4.48)
Note that this result does not depend on the momentum p in any way.
The integral is divergent for d 2 which is consistent with the formula for the degree of
divergence derived above; this is reected in that (1
d
2
) has poles at d = 2 and d = 4. To
exhibit these we use the recurrence relation for Gamma functions to obtain
(x 1) =
(x + 1)
x(1 x)
=
1
x
e
x
_
1 +x +O(x
2
)
_
, (4.49)
in the limit x 0. Now let us consider
d = 4 , (4.50)
where in due course we will take the limit 0. Then we can use this expansion to rewrite

_
1
d
2
_
=
_

2
1
_
=
2

1
2

_
1 +

2
+O(
2
)
_
, (4.51)
in (4.48), which gives the asymptotic expression for as 0

(1)
2
(p, p)
1
2

_
2

+ 1
_
(m
2
)
d
2
1
(4)
d
2
e

1
2

1
2

16
2
m
2
_
2

+ 1 log m
2
_
[e

4]
1
2

. (4.52)
In this result the divergent part, as a pole in , has been separated from the nite part. This
is true generally using the method of dimensional regularisation for Feynman integrals; the
divergencies which may be present are reduced to poles in .
A rather more non-trivial one loop example corresponds to the graph
53
-
p p

_ -

k
k +p
g g
for which the associated Feynman integral is

(1)
2
(p, p) =
1
2
g
2
_
d
d
k
(2)
d
1
(k
2
+m
2
)((k +p)
2
+m
2
)
. (4.53)
To evaluate this we use
1
x
=

_
0
d e
x
(4.54)
to rewrite the expression as

(1)
2
(p, p) =
1
2
g
2

_
0
d
1

_
0
d
2
_
d
d
k
(2)
d
e

1
(k
2
+m
2
)
2
((k+p)
2
+m
2
)
. (4.55)
The exponent is then simplied by completing the square,

_
(
1
+
2
)k
2
+

1

1
+
2
p
2
+ (
1
+
2
)m
2
_
, k

= k +

2

1
+
2
p . (4.56)
Now we can shift the integral d
d
k d
d
k

; there are no surface terms when d such that the

integral is nite. The basic integral we need becomes
_
d
d
k

(2)
d
e
(
1
+
2
)k
2
=
1
(4)
d
2
1
(
1
+
2
)
d
2
, (4.57)
and using this

(1)
2
(p, p) =
1
2
g
2
1
(4)
d
2

_
0
d
1

_
0
d
2
1
(
1
+
2
)
d
2
e

1
+
2
p
2
(
1
+
2
)m
2
. (4.58)
The divergences which are present when d = 4 arise from the singular behaviour of the integrand
as
1
,
2
0. These may be exhibited explicitly, and the result reduced to a single integral,
by making the substitution

1
+
2
= s ,
1
= s,
2
= s(1 ) , (4.59)
where 0 < s < and 0 < < 1, and
d
1
d
2
= sds d. (4.60)
Hence we obtain

(1)
2
(p, p) =
1
2
g
2
1
(4)
d
2
1
_
0
d

_
0
ds s
1
d
2
e
s((1)p
2
+m
2
)
. (4.61)
The s integral gives a Gamma function so the result becomes

(1)
2
(p, p) =
1
2
g
2
1
(4)
d
2

_
2
d
2
_
1
_
0
d
_
(1 )p
2
+m
2
_d
2
2
. (4.62)
For d = 4 , we want to nd the leading terms as 0, if we are interested in the theory
in four dimensions. We use the asymptotic form for x 0
(x) =
1
x
e
x
_
1 +O(x
2
)
_
, (4.63)
54
to obtain

(1)
2
(p, p)
1
2
g
2
16
2
_
_
2

1
_
0
d log
_
(1 )p
2
+m
2
_
_
_
[e

4]
1
2

. (4.64)
Again this is expressed in terms of a divergent pole term and a nite part. In this case the
integral depends on the momentum, but the divergent part does not.
35
Now consider n = 4, a
4
interaction and only one-loop graphs:
@
@R
@
@

-
k
k +p

@
@I
@
@
+
p
1
p
2
p
3
p
4
@
@R
@
@

p
1
p
3
`
_

@
@I
@
@
p
2
p
4
+
@
@R
@
@

p
1
p
2
`
_ @
@
@
@I
p
4
@

p
3
In this case we need to consider

(1)
4
(p
1
, p
2
, p
3
, p
4
) =

P=p
1
+p
2
, p
1
+p
3
, p
1
+p
4
1
2

2
_
d
d
k
(2)
d
1
(k
2
+m
2
)((k +P)
2
+m
2
)
, (4.65)
where the sum corresponds to the three one-loop graph contributions. The divergent part of
the integral follows from previous calculations,

(1)
4
(p
1
, p
2
, p
3
, p
4
)
3
2

2
16
2
2

. (4.66)
Other graphs are nite, for example
@
@

@
@

@
@
@
while these graphs with one three- and one four-point vertex also have a divergence
@
@
@

_
+
@
@

`
_
+

@
@
`
_
The divergent term will be of the order
3
2
g
16
2
2

. (4.67)
There is also a one loop graph with one external line
-
p = 0

_
35
One should be concerned that we take the logarithm of a dimensionful quantity; we will come back to that
later.
55
and the associated integral gives

(1)
1
(0) =
1
2
g
_
d
d
k
(2)
d
1
k
2
+m
2

gm
2
16
2

. (4.68)
Let us summarise the above one-loop results for the n-point functions
(1)
n
(p
1
, . . . , p
n
), ne-
glecting the nite parts:

(1)
1
(0)
1
16
2

gm
2
, (4.69a)

(1)
2
(p, p)
1
16
2

(m
2
+g
2
) , (4.69b)

(1)
3
(p
1
, p
2
, p
3
)
1
16
2

3g, (4.69c)

(1)
4
(p
1
, p
2
, p
3
, p
4
)
1
16
2

3
2
. (4.69d)
If we had introduced a cuto instead of dimensional regularisation then we would eectively
replace
1

by
1
2
log
2
.
The essential claim is, in a quantum eld theory arising from a Lagrangian L, these
divergent contributions can be cancelled by adding new terms to the Lagrangian. When you
rst meet this, it may seem rather arbitrary, ad hoc and perhaps unremarkable, but there is
a deep signicance in that this can be carried out consistently to all orders of perturbation
theory. The additional terms added to L are termed counterterms and have the form
L
c.t.
= A
1
2

V
c.t.
() , (4.70)
where V
c.t.
is a local potential expressible in this theory as
V
c.t.
() = E +
1
2
B
2
+
1
6
C
3
+
1
24
D
4
. (4.71)
For each term in L
c.t.
there are extra vertices which generate additional contributions. For
those corresponding to tree-level diagrams we have

(0)
2
(p, p)
c.t.
= Ap
2
B,
(0)
4
(p
1
, p
2
, p
3
, p
4
)
c.t.
= D,

(0)
1
(0)
c.t.
= E ,
(0)
3
(p
1
, p
2
, p
3
)
c.t.
= C . (4.72)
,
@

@
@
@
@
@
@

We choose the coecients to cancel the divergences. At this order,

A
(1)
= 0 , B
(1)
=
1
16
2

(m
2
+g
2
) , C
(1)
=
1
16
2

3g, D
(1)
=
1
16
2

3
2
, E
(1)
=
1
16
2

gm
2
.
(4.73)
Then
(1)
n
(p
1
, . . . , p
n
) +
(0)
n
(p
1
, . . . , p
n
)
c.t,
has no pole in as 0 and we may set = 0 with
a nite result.
More succinctly, all contributions to the one-loop counterterm potential can be written as
V
c.t.
()
(1)
=
1
32
2

()
2
. (4.74)
To check that that is the case
V

() = m
2
+g +
1
2

2
, (4.75)
V

()
2
= m
4
+ 2m
2
g + (g
2
+m
2
)
2
+g
3
+
1
4

4
, (4.76)
56
and the coecients in V
c.t.
()
(1)
of ,
1
2

2
,
1
6

3
,
1
24

4
match with E
(1)
, B
(1)
, C
(1)
, D
(1)
above.
The Lagrangian L +L
c.t.
then ensures that all one-loop graphs are nite.
Let us make some remarks:
(i) B, C, D and E are arbitrary to the extent that nite pieces can be added, but keeping
just the poles in gives a unique prescription. This is in a way arbitrary, but sucient.
(ii) If V () is a polynomial of degree four in , then so is V

()
2
. As will be apparent later
this is crucial in ensuring renormalisability of this theory.
4.2.2 Dimensional Regularisation for Two-Loop Graphs
To show how divergencies can be consistently removed we need to proceed to higher numbers
of loops, L 2. The result that Feynman integrals can be regularised in a fashion consistent
with the general requirements of quantum eld theory is no longer almost a triviality but
requires a careful analysis of the divergencies and also sub-divergencies in Feynman integrals.
In terms of calculations, one-loop diagrams are fairly easy, sometimes calculating two loops
is also straightforward, but just requires more work. Of course when quantum eld theory
calculations were rst done in the 1940s, even one loop seemed to be hard, but now we know
many techniques of how to do this.
Consider now
(2)
2
(p, p) with just a
4
interaction V () =
1
24

4
. There are basically two
two-loop diagrams:
(a) (b)
-
_
_
?
6
?
6
p
k k
l l
-

_
p
The contribution of the rst diagram, after Wick rotation, is given by

(2)
2
(p, p)
a
=
1
4

2
_
d
d
k
(2)
d
1
(k
2
+m
2
)
2
_
d
d
l
(2)
d
1
l
2
+m
2
. (4.77)
The two momentum integrals are independent so we can use
_
d
d
l
(2)
d
1
l
2
+m
2
=

_
0
d
_
d
d
l
(2)
d
e
(l
2
+m
2
)
=
1
(4)
d
2

_
0
d e
m
2 1

d
2
=
1
(4)
d
2

_
1
d
2
_
(m
2
)
d
2
1
, (4.78)
as previously, and also
_
d
d
k
(2)
d
1
(k
2
+m
2
)
2
=

m
2
_
d
d
k
(2)
d
1
k
2
+m
2
=
1
(4)
d
2

_
2
d
2
_
(m
2
)
d
2
2
. (4.79)
Both are expressible in terms of Gamma functions and hence we have the result

(2)
2
(p, p)
a
=
1
4

2
1
(4)
d

_
2
d
2
_

_
1
d
2
_
(m
2
)
d3
=
1
4

2
1
(4)
d

_
2
d
2
_
2
1
d
2
(m
2
)
d3
(4.80)
57
Setting d = 4 and using (

2
) e

(
2

+O()) we obtain

(2)
2
(p, p)
a


2
4(4)
4
m
2(1)
_
4

2
+
2

_
[e

4]

=

2
(4)
4
m
2
_
1

2
+
1
2

1

log m
2
_
[e

4]

. (4.81)
This has a double pole at = 0.
Since the initial Lagrangian has been modied by the additional term L
c.t.
which generates
the counterterms necessary to subtract the divergent pole terms for one loop Feynman integrals
we must also consider Feynman graphs involving vertices generated by this. With g = 0 at this
order we need consider only
1
2
B
2
+
1
24
D
4
. (4.82)
For one loop graphs with two external lines and to rst order in B, D we have

_
B

_
D
The corresponding integrals are

(1)
2
(p, p)
c.t.
=
1
2
B
(1)

_
d
d
k
(2)
d
1
(k
2
+m
2
)
2

1
2
D
(1)
_
d
d
k
(2)
d
1
k
2
+m
2
. (4.83)
Let us write D
(1)
=
1
3
D
(1)
+
2
3
D
(1)
and for the moment keep only the
1
3
D
(1)
term (the re-
maining
2
3
D
(1)
will contribute elsewhere). The one-loop contributions from these counterterms,
with the previously calculated results for B
(1)
, D
(1)
, are then

(1)
2
(p, p)
c.t.,a
=
1
2
m
2
16
2

_
d
d
k
(2)
d
1
(k
2
+m
2
)
2

1
2

2
16
2

_
d
d
k
(2)
d
1
k
2
+m
2
=
1
2

2
m
2
16
2

(m
2
)

2
(4)
d
2

_
2
d
2
_
_
1
1
1
d
2
_

2
m
2
(4)
4
_
2

2
+
1
2

1

log m
2
_
[e

4]
1
2

. (4.84)
Taking everything together we get

(2)
2
(p, p)
a
+
(1)
2
(p, p)
c.t.,a


2
m
2
(4)
4
1

2
+ nite terms . (4.85)
In this result there are no
1

log m
2
terms - the residue is polynomial in and m
2
.
The calculation for the second graph is more complicated, so we will only sketch this:
- -

_
-
p p l
-
-
k
p k l
The corresponding Feynman integral is

(2)
2
(p, p)
b
=
1
6

2
_
d
d
k
(2)
d
d
d
l
(2)
d
1
(k
2
+m
2
)(l
2
+m
2
)((p k l)
2
+m
2
)
. (4.86)
58
There are sub-divergences as k for nite l, as l for nite k and as k for nite
k +l.
There are various ways in which we can treat this integral; to work this out in full is pretty
complicated. First consider the l integral:
_
d
d
l
(2)
d
1
(l
2
+m
2
)((p

l)
2
+m
2
)
=
(2
d
2
)
(4)
d
2
1
_
0
d
_
(1 )p
2
+m
2
_d
2
2
, p

= p k .
(4.87)
If this is then inserted into the k-integral the result is a bit messy. We therefore consider the
special case m
2
= 0, which allows the integrals to be done completely. Then
_
d
d
l
(2)
d
1
l
2
(p

l)
2
=
(2
d
2
)
(4)
d
2
K
d
(p
2
)
d
2
2
, K
d
=
1
_
0
d ((1 ))
d
2
2
. (4.88)
Now substitute this into the initial integral to get

(2)
2
(p, p)
b

m
2
=0
=
1
6

_
2
d
2
_
(4)
d
2
K
d
_
d
d
k
(2)
d
1
k
2
((k p)
2
)
d
2
2
. (4.89)
To rewrite this we may use the standard representations
1
k
2
=

_
0
d
2
e

2
k
2
;
_
2
d
2
__
(k p)
2
_d
2
2
=

_
0
d
1

1
d
2
1
e

1
(pk)
2
, (4.90)
and then this implies

(2)
2
(p, p)
b

m
2
=0
=
1
6

2
1
(4)
d
2
K
d

_
0
d
1

_
0
d
2
_
d
d
k
(2)
d
e

2
k
2

1
d
2
1
e

1
(pk)
2
, (4.91)
so that we can complete the square

2
k
2
+
1
(p k)
2
= (
1
+
2
)k

2
+

1

1
+
2
p
2
, k

= k

1

1
+
2
p . (4.92)
It is now easy to carry out the k

-integration:

(2)
2
(p, p)
b

m
2
=0
=
1
6

2
(4)
d
K
d

_
0
d
1

_
0
d
2

1
d
2
1
(
1
+
2
)
d
2
e

1
+
2
p
2
. (4.93)
Now substitute
1
+
2
= s,
1
= s,
2
= (1 s) to obtain

(2)
2
(p, p)
b

m
2
=0
=
1
6

2
(4)
d
K
d
1
_
0
d
1
d
2

_
0
ds s
2d
e
s(1)p
2
=
1
6

2
(4)
d
K
d
(3 d) (p
2
)
d3
1
_
0
d
d
2
2
(1 )
d3

1
12

2
(16
2
)
2

p
2
, (4.94)
noting that, with d = 4 , (3 d)
1

, K
4
= 1. This can be cancelled with a contribution
A
(2)
=
1
12

2
(16
2
)
2

. (4.95)
59
Another special case which is tractable is to consider m
2
,= 0, but p
2
= 0. Following the
same approach as before

(2)
2
(0, 0)
b
=
1
6

_
2
d
2
_
(4)
d
2
_
d
d
k
(2)
d
1
k
2
+m
2
_
1
0
d
_
(1 )k
2
+m
2
)
d
2
2
. (4.96)
Writing both factors as exponentials, with integrals over
1
,
2
, allows the k-integration to be
carried out. The
1
,
2
integrations can be treated in the same fashion as previously giving

(2)
2
(0, 0)
b
=
1
6

2
(m
2
)
d3
1
(4)
d
(3 d) J
2
, (4.97)
for
J
2
=
_
1
0
d
_
1
0
d
1
d
2
_
1 +(1 )
_

d
2
. (4.98)
At this point the integral is no longer in the domain of opening theory, if this were chess, for
evaluating Feynman integrals but it can be carried out in terms of hypergeometric functions.
What is required here however is just to determine the poles in as 0. There is a pole
due to the (3 d) factor but there are also poles coming from J
2
. It is not dicult to see
that there is a pole arising from the divergence of the -integration as 0 but there are also
poles from divergences at 1 and 0, 1. These correspond to the sub-divergencies present
in the original Feynman integral. To disentangle these it is convenient to introduce into the
integral expression for J
2
1 = (1 ) + +(1 ) . (4.99)
Each of the three terms which arise are identical. This reects the symmetry of the original
graph under permutation of lines but can also be demonstrated by considering a change of
variables

= + 1 ,

= so that (1 +(1 )) =

(1

(1

))
and dd =

. Hence we can write

J
2
= 3
_
1
0
d
_
1
0
d
1
d
2
(1 )
_
1 +(1 )
_

d
2
. (4.100)
This expression now has a divergence only when 0 so that J
2
can be decomposed
J
2
= 3
_
1
0
d
_
1
0
d
1
d
2
+J
2

= 3
2

+J
2

,
J
2

= 3
_
1
0
d
_
1
0
d
1
d
2
_
(1 )
_
1 +(1 )
_

d
2
1
_
, (4.101)
where J
2

is nite for d = 4 since there is no divergence at = 0. Hence letting 1/

J
2

d=4
= 3
_
1
0
d
_

1
d
_
1
_
1 +(1 )
_
2

1

_
= 3
_
1
0
d
_
log (1 ) 1
_
= 3 . (4.102)
The expression for J
2
then becomes
J
2
= 3
_
2

+ 1
_
+ O() . (4.103)
With this result for J
2
and applying (4.49) for (3 d) = ( 1) we then have in (4.97)

(2)
2
(0, 0)
b
=
1
2

2
m
2
(16
2
)
2
_
2

2
+
3

log m
2
_
[e

4]

+ nite parts . (4.104)

There is also additional corresponding contribution comes from the the counterterms where
we consider the contribution from
2
3
D
(1)
(since we took account of the
1
3
D
(1)
part previously)
60
which is given by

(2)
2
(0, 0)
c.t.,b
=
1
2
2
2
16
2

_
d
d
k
(2)
d
1
k
2
+m
2
=
1
2
2
2
16
2

1
(4)
d
2

_
1
d
2
_
(m
2
)
d
2
1

2
(16
2
)
2
m
2
_
2

2
+
1

log m
2
_
[e

4]
1
2

+ nite parts , (4.105)

where the integral was evaluated before. Adding this to
(2)
2
(0, 0)
b
gives

(2)
2
(0, 0)
b
+
(2)
2
(0, 0)
c.t.,b


2
m
2
(16
2
)
2
_
1

2

1
2
_
, (4.106)
and again there is no log m
2
term left.
If we combine both two loop graphs by adding the results in (4.85), (4.94) and (4.106) the
two loop divergent part becomes

(2)
2
(p, p) +
(2)
2
(p, p)
c.t.

1
12

2
(16
2
)
2

p
2
+

2
m
2
(16
2
)
2
_
2

2

1
2
_
. (4.107)
This is the general result as contributions to the Feynman integrals which have been neglected
are all nite when d = 4. Hence the divergences which are present in
(2)
2
(p, p), after sub-
tracting sub-divergencies as shown above, can now be cancelled by additional contributions to
A and B, which give a contribution to
2
of the form Ap
2
B. The two loop results, ensuring
that
(2)
2
is nite, are then determined to be
A
(2)
=
1
12

2
(16
2
)
2

, B
(2)
=

2
m
2
(16
2
)
2
_
2

2

1
2
_
, (4.108)
which involve both double and single poles in .
4.3 Renormalisation and the Renormalisation Group
The results of one and two loop calculations can be generalised to the crucial result for pertur-
bative quantum eld theory:
Renormalisation Theorem
If all sub-divergences in a Feynman integral are subtracted, then depending on the degree of
divergence D,
(i) if D < 0, then the integral is nite,
(ii) if D 0, then the divergent part (i.e. the poles in ) is proportional to a polynomial in the
external momenta and the couplings of dimension D.
(iii) the divergences are cancelled by appropriate counterterms which determine L
c.t.
, these
counterterms then generate additional Feynman graphs whose corresponding Feynman integrals
then cancel sub-divergences in higher loop integrals.
For the
4
theory then
4
has D = 0 and the divergent part is a dimensionless constant
depending only on the coupling while
2
has D = 2 and the divergent part is linear in m
2
and p
2
with coecients depending on . For a general potential V () which is a polynomial of
degree four,
D 4 E . (4.109)
We then only need counterterms of degree four in . So we may restrict L
c.t.
to the form
specied by A and V
c.t.
(). In fact, we can write
A =

n1
a
n
()

n
, V
c.t.
() =

n1
1

n
V
n
() . (4.110)
V
n
() is a polynomial which depends on V

(), V

() and , and is of degree four in , e.g.

at one loop
A
(1)
= 0 , V
(1)
c.t.
() =
1
32
2

()
2
. (4.111)
61
The basic theorem is true for any quantum eld theory. However, for non-renormalisable
theories, D increases with the number of loops and is not restricted by the number of external
lines. In this situation, the countertermL
c.t.
becomes arbitrarily complex with many parameters
as the order of the calculation increases. For renormalisable theories, D is bounded; L
c.t.
may
be restricted to the same form as the original Lagrangian.
Ultimately the theorem demonstrates that the limit
Z[J]
ren.
= lim
0
_
d[] e
i(S[]+S
c.t.
[])+i
R
d
d
x J(x)(x)
, (4.112)
exists to any order of perturbation theory so long as S
c.t.
is chosen suitably. As a consequence
we have nite renormalised connected 1PI functions
n
(p
1
, . . . , p
n
)
ren.
.
4.3.1 Bare Lagrangians
Let us now restrict our attention to scalar
4
theory. We know on the basis of the renormali-
sation theorem, as described above, that adding
L
c.t.
= A
1
2

V
c.t.
() (4.113)
to the initial Lagrangian L is sucient to all orders of perturbation theory, for suitable A and
V
c.t.
() which is itself a polynomial of degree four, to ensure that the resulting theory determines
nite correlation functions of arbitrary numbers of scalar elds (x
1
) . . . (x
n
)). Let us dene
a bare Lagrangian by
L
0
= L +L
c.t.
= Z
1
2

V () V
c.t.
() , (4.114)
where
Z = 1 +A. (4.115)
This can be simplied by a rescaling of to ensure the coecient of the kinetic term is the
same as in the initial theory, thus

0
=

Z, (4.116)
and hence then
L
0
=
1
2

0
V
0
(
0
) . (4.117)
Since A is given by a formal power series in then

Z may also be dened as a power series
in without ambiguity. It is essential that Z > 0 but this is never an issue in perturbation
theory. We also call
0
the bare eld and V
0
(
0
) the bare potential which is again a quartic
polynomial.
Now suppose that g = 0, so that
V
c.t.
() =
1
2
B
2
+
1
24
D
4
; V
0
(
0
) =
1
2
m
2
0

2
0
+
1
24

4
0
, (4.118)
where

0
=
+D
Z
2
, m
2
0
=
m
2
+B
Z
. (4.119)
From the one-loop results obtained so far
m
2
0
= m
2
_
1 +

16
2

_
,
0
=
_
1 +
3
16
2

_
, Z = 1 +O(
2
) . (4.120)
The so-called bare quantities m
2
0
and
0
are expressible in terms of the original parameters as
a power series in with more and more divergent terms or higher and higher poles in .
The functional integral may be rewritten in terms of
0
and L
0
:
S
0
[
0
] =
_
d
d
x L
0
(
0
), d[] = Nd[
0
], (4.121)
62
where N is some constant, so that

F(
0
)
_
=
1
Z
0
_
d[
0
] e
iS
0
[
0
]
F(
0
) , (4.122)
with Z
0
chosen so that 1) = 1.
However, we avoided one slight subtlety which we now have to take into account, related
with a dimensional analysis; in the process of regularisation it is necessary to introduce a mass
scale. A cut-o introduces an obvious mass scale. In dimensional regularisation however it
has not been perhaps apparent, and is slightly more subtle. We extended L to d ,= 4. This is
something one should worry about, because dimensions change as you change dimension. Since
an action has dimension zero, L must have dimension d, so that V () also has dimension d,
whereas has dimension
1
2
d 1. These are ordinary, everyday dimensions. For a potential
V () =
1
2
m
2

2
+
1
6
g
3
+
1
24

4
, (4.123)
this means that m
2
has dimension 2 as usual while g has dimension 3
1
2
d and has dimension
4 d. In the end the dimension should be four and the physical couplings g and , which
parameterise the theory, should have dimension 1 and 0, as they would in other regularisation
schemes. This can be achieved by replacing in V () in dimensions d ,= 4
g g

2
,

, (4.124)
where denes an (arbitrary) mass scale. Hence with d = 4 we now write the potential
term as
V () =
1
2
m
2

2
+
1
6
g

3
+
1
24

4
. (4.125)
Correspondingly in the counterterm potential we let D

D and C
1
2

C.
4.3.2 The Callan-Symanzik Equation
For simplicity we consider the case g = 0 when the bare Lagrangian is
L
0
=
1
2

1
2
m
2
0

2
0

1
24

4
. (4.126)
This depends on m
2
0
,
0
and
0
and clearlyL
0
is then independent of .
The bare quantities are functions of the nite , m
2
. With the prescription that the coun-
terterms necessary to ensure niteness involve contributions which are just poles in we have
relation of the form

0
= +F

(, ) , F

(, ) =

n=1
f
n
()

n
, (4.127)
and also
m
2
0
= m
2
_
1 +F
m
2(, )
_
, F
m
2(, ) =

n=1
b
n
()

n
. (4.128)
Similarly, we can write that
Z = 1 +

n1
a
n
()

n
. (4.129)
The nite physical correlation functions depend then on , m
2
and also which appears
to be more than the bare theory. In fact is arbitrary but to show this more precisely it is
necessary to consider the relation between the bare theory results, which are independent of ,
and the corresponding nite quantities obtained by the regularisation procedure. Correlation
functions for the bare elds may be dened formally by a functional integral,
G
0
n
(x;
0
, m
2
0
) :=

0
(x
1
) . . .
0
(x
n
)
_
=
1
_
d[
0
] e
iS
0
[
0
]
_
d[
0
]
0
(x
1
) . . .
0
(x
n
) e
iS
0
[
0
]
,
(4.130)
63
where x = (x
1
, . . . , x
n
) and the result depends on the bare parameters
0
and m
2
0
. Since

0
=

Z
G
n
(x; , m
2
, ) :=

(x
1
) . . . (x
n
)
_
= Z

n
2
G
0
n
(x;
0
, m
2
0
) , (4.131)
and crucially G
n
(x; , m
2
, ) is a nite function of , m
2
, with a non singular limit as 0.
The critical observation is that this also depends on since this is a necessary part of the
regularisation procedure. Because bare quantities do not depend on the mass scale , it is clear
that

d
d
G
0
n
(x;
0
, m
2
0
) = 0 , (4.132)
and it follows that

d
d
_
Z
n
2
G
n
(x; , m
2
, )
_

0
,m
2
0
= 0 . (4.133)
By using the standard chain rule this becomes a linear rst order partial dierential equation.
First dene

:=
d
d

0
,
m
2 :=
d
d
m
2

m
2
0
,
0
,

Z :=
d
d

0
. (4.134)
Then we obtain
_

+
m
2

m
2
+n

_
G
n
(x; , m
2
, ) = 0 . (4.135)
This equation will play an important role; it is called the Callan-Symanzik
36
or renormalisa-
tion group (RG) equation. Since G
n
is nite, i.e. there are no poles in , the quantities

,
m
2
and

must also be nite and have no poles in . The RG equation follows from perturbation
theory and essentially the renormalisation theorem stated above. Similar equations are valid
for any renormalisable quantum eld theory and are assumed to be an exact property.
Let us illustrate how we can determine the functions appearing in the RG equation. By
considering the relation between
0
and we have

d
d
(

0
)

0
=

0
= ( +F

)
=
d
d
_
+F(, )
_

0
=

_
+F(, )
_
=

. (4.136)
If we dene

by

= +

, (4.137)
then this can be re-expressed as

= F

( +

=
_

1
_
F

. (4.138)
Now since F

is given by an expansion in
n
, for n = 1, 2, . . . , and since

has no poles in
the equation is only consistent if

() =
_

1
_
f
1
() , (4.139)
and in fact

is then independent of . Furthermore the pole terms,

n
with n > 0, on the
right hand side must cancel which gives
_

1
_
f
n+1
() =

()

f
n
() . (4.140)
This recursive relation enables us to calculate f
n
() ultimately in terms of f
1
() and

()
which is also determined by f
1
().
36
Callan, Curtis (1942-); Symanzik, Kurt (1923-1983)
64
We can obtain similar results from the relation between m
2
0
and m
2
, . Taking derivatives
with respect to :
0 =
d
d
m
2
0

m
2
0
,
0
=
d
d
_
m
2
(1 +F
m
2)
_

m
2
0
,
0
=
_

+
m
2

m
2
_
_
m
2
(1 +F
m
2)
_
=
m
2(1 +F
m
2) +m
2

F
m
2 . (4.141)
Dening a new function
m
2() by

m
2 = m
2

m
2() , (4.142)
this gives

m
2 = ( +

F
m
2
m
2F
m
2 . (4.143)
By the same argument as before,
m
2 is nite and so has to be independent of . It is determined
just by the simple poles in F
m
2

m
2() =

b
1
() , (4.144)
and also the higher order poles are iteratively determined by

b
n+1
() =
_

m
2() +

()

_
b
n
() . (4.145)
The denition of

may be rewritten
Z

=
1
2

d
d
Z

0
. (4.146)
Inserting Z = 1 +A, this gives
(1 +A)

=
1
2

d
d
A

0
=
1
2

A, (4.147)
which implies

=
1
2
( +

A. (4.148)
Looking at the O(
0
) coecients gives

() =
1
2

a
1
() , (4.149)
and also to O(
n
)

a
n+1
() =
_

()

()
_
a
n
() . (4.150)
Let us consider what happens for one loop for this scalar theory. The results that we
previously derived were

0
= +
3
2
16
2

+O(
3
) , (4.151a)
m
2
0
= m
2
_
1 +

16
2

+O(
2
)
_
, (4.151b)
which gives to lowest order

() =
3
2
16
2
+O(
3
) , (4.152a)

m
2() =

16
2
+O(
2
) , (4.152b)
and we also have

() = O(
2
).
65
4.3.3 Evolution of Coupling Constants
We consider here solving the Callan-Symanzik equation but since this may be derived in any
renormalisable theory we consider this in general without restriction to any particular theory.
For simplicity we consider a single dimensionless coupling g with a corresponding -function
(g). If we set any mass m = 0 the equation has the generic form
_

+(g)

g
+(g)
_
G(p; g, ) = 0 , (4.153)
for G(p; g, ) a physical function of a set momenta p and also of the coupling g as well as
the regularisation scale .
If G is dimensionless, it can only depend on the quotient
p

for all momenta:

G(p; g, ) = F
__
p

_
; g
_
, (4.154)
for some function F. Thus we can regard G as only a function of g and . It is always possible
to ensure that G is dimensionless by factoring out suitable momenta. The crucial point is then
that the dependence on p, which is physically interesting, is related to the dependence on the
arbitrary mass scale . For the moment we drop and be restored later. The basic equation
is then simply
_

+(g)

g
_
G(g, ) = 0 , (4.155)
where the dependence on the momenta is left implicit. There exists a standard procedure, the
method of characteristics, to solve this; dene a quantity g() (which is called the running
coupling) by the requirement that it is a solution of

d
d
g = (g) . (4.156)
This equation can be solved by
g()
_
g(
0
)
dg
(g)
= log

0
. (4.157)
Then the partial dierential equation is reduced to

d
d
G(g(), ) = 0 , (4.158)
which requires that G(g(), ) is independent of or
G(g(), ) = G(g(
0
),
0
) . (4.159)
This is a reection of the arbitrariness of , the direct dependence on is compensated by the
dependence on g(), so that , g are replaced by a single parameter.
If there is only one momentum, so that F = F(
p
2

2
; g), the solution shows that
f
_
p
2

2
; g()
_
= F
_
p
2

2
0
; g(
0
)
_
. (4.160)
We may choose
0
= p and then
f
_
p
2

2
; g()
_
= F (1; g(p)) , (4.161)
so that the dependence on p is just given by g(p). This allows us to make statements about the
properties of the theory which follow from analysing how it depends on the coupling.
First we consider the qualitative features the solution for the running coupling g(). If
(g) > 0 then g increases with , whereas if (g) < 0 then g decreases with . A graph for
(g) if (g) > 0 for small g and also if (g

) = 0 for some nite g

has the form

66
6
- - ,
g

g
(g)
For such a functional dependence g() g

as . We can say that g

is an ultraviolet
(UV) xed point. If we assume (g) is negative for small g, the picture looks like this:
6
- ,
g
(g)
In this case g() 0 as . (Note that always (0) = 0.) In this case g = 0 is an
ultraviolet xed point. This situation is referred to as asymptotic freedom.
Let us now bring back into the equation and see what modication of the solution is
required. For
_

+(g)

g
+(g)
_
G(g, ) = 0 , (4.162)
we introduce g() as before and the equation becomes

d
d
G(g(), ) = (g()) G(g(), ) . (4.163)
This can be integrated to give
G(g(), ) = e

0
ds
s
(g(s))
G(g(
0
),
0
) . (4.164)
The exponent may also be rewritten by a change of integration variable as

0
ds
s
(g(s)) =
g()
_
g(
0
)
dg
(g)
(g)
. (4.165)
If there is a UV xed point g

then

0
ds
s
(g(s)) (g

) log as . (4.166)
For the function of a single momentum F these results give
f
_
p
2

2
; g()
_
= e
R
p

ds
s
(g(s))
F
_
1; g(p)
_
. (4.167)
Asymptotically, when there is an ultraviolet xed point, as p ,
f
_
p
2

2
; g()
_
p
(g

)
F(1; g

) . (4.168)
67
When g is small we may use perturbative results. As an example suppose
(g) = bg
3
. (4.169)
If b > 0 there is asymptotic freedom and as g() 0 for large the small g perturbative results
become a valid approximation. The running coupling is then given by the following integral,

1
b
g()
_
g(
0
)
dg
g
3
= log

1
g()
2

1
g(
0
)
2
= 2b log

0
. (4.170)
Now the right-hand side goes to innity as , and so g()
2
0 if b > 0. We can rewrite
this as
1
g()
2
= b log

2

2
, (4.171)
for some constant . Alternatively the constant , which appears as a constant of integration,
is given by
= e

1
2bg()
2
. (4.172)
Let us suppose that (g) = cg
2
as well; then the integral in the solution of the RG equation
gives
g(p)
_
g()
dg
(g)
(g)
=
c
b
log
g(p)
g()
. (4.173)
The solution for F(p
2
/
2
; g) becomes
f
_
p
2

2
; g()
_
=
_
g
2
(p)
g
2
()
_

c
2b
F
_
1; g(p)
_
. (4.174)
As p , g(p) 0, with asymptotic freedom, so that perturbative results for (g), (g)
and also F(p
2
/
2
; g) can be used to give a precise, well justied, prediction for this limit. This
illustrates how asymptotic freedom can be use to nd out about the behaviour of quantum eld
theories such as QCD for large momenta when perturbative results show (g) < 0.
As a nal consideration, let us now introduce a mass term, assuming that we have an
equation (here for a single coupling g)
_

+(g)

g
+
m
2(g)m
2

m
2
+(g)
_
G(g, m
2
, ) = 0 . (4.175)
As before we solve this by introducing a running coupling g() and also a running mass m(),
which is determined by the analogous equation

d
d
m
2
=
m
2(g())m
2
. (4.176)
This can be solved by
m
2
() = m
2
(
0
) e
R

0
ds
s

m
2(g(s))
. (4.177)
constant g. The solution of the RG equation now becomes
G
_
g(), m
2
(),
_
= e

0
ds
s
(g(s))
G
_
g(
0
), m
2
(
0
),
0
_
, (4.178)
which reduces to the previous solution when m
2
= 0. For a function of a single momentum p
we let G F(p
2
/
2
, m
2
/
2
, g) and the solution becomes
f
_
p
2

2
,
m
2

2
, g()
_
= e
R
p

ds
s
(g(s))
F
_
1,
m
2
(p)
p
2
, g(p)
_
. (4.179)
68
Now again consider the situation where we take the limit p and we have an ultraviolet
xed point g

, so that g(p) g

for p . In this case

0
ds
s

m
2(g(s))
m
2(g

) log as . (4.180)
So long as
m
2(g

) < 2 then the dependence on m

2
can be neglected in the asymptotic limit.
For asymptotically free theories
m
2(g) = O(g
2
) and the masses can also be neglected in the
UV limit.
4.4 Eective Potentials
In classical dynamics, if there is a potential V (), then the ground state is determined by the
minimum of V . If V (
0
) = V
min.
then in the ground state =
0
. This may be degenerate or
there may be several minima of V () each of which dene potential stable ground states. Since
in a quantum eld theory the classical potential requires additional counterterms it is necessary
to reconsider how the ground state is determined.
In quantum eld theory, what we mean by the ground state is the vacuum [0), which is
required to be the state of zero energy. The quantum ground state is determined by a modied
eective potential V
e.
(), which we show here how to dene and also how to calculate in
perturbative quantum eld theory. The essential denition is through the generating functional
[] for connected one-particle irreducible graphs. We have
[]

=const.
= 1 V
e.
() , 1 =
_
d
d
x, (4.181)
where 1 is the volume of spacetime (for this to be well dened it is necessary to consider a nite
region and then take the limit as the volume becomes large). V
e.
() is equal to the minimum
energy density subject 0

[0

) = . The quantum vacuum is then [0

0
) with
0
determined
by the global minimum of V
e.
().
In perturbation theory V
e.
() may be calculated as an expansion with the rst term the
classical potential V (); we will explain in outline how we can do that, at least in the simplest
case, and also show how divergencies are treated in this case. We discuss only the lowest
approximation beyond the classical potential, but this can be extended to higher orders. The
basic denitions are
e
iW[J]
=
_
d[] e
iS[]+i
R
d
d
x J(x)(x)
, [] = W[J] +
_
d
d
xJ(x)(x) , (4.182)
where
W[J]
J(x)
= (x) = (x))
J
. (4.183)
Here . )
J
denotes that it is calculated with the action S
J
[] = S[] +
_
d
d
x J(x)(x). The
perturbation expansion is derived by rst taking (x) = +f(x), with here a constant and
is sometimes referred to as the background eld. The action is then expanded in terms of f(x)
to obtain
S[] = S[] +
_
d
d
x
S[]
(x)
f(x)
1
2
_
d
d
x f(x)f(x) +O(f
3
) . (4.184)
For the simple scalar eld theory with the action
S[] =
_
d
d
x
_

1
2
V ()
_
, (4.185)
we have
S[]
(x)
:= J

=
2
V

() = V

() , for const. , (4.186)

and

_
d
d
x f(x)f(x) =
_
d
d
x
_
f f V

()f
2
_
, =
2
+V

() . (4.187)
69
Taking J(x) = J

to cancel the O(f) terms and neglecting the O(f

3
) terms in the expansion
of the action and using d[] = d[f] gives
e
iW[J

]
e
iS[]+i
R
d
d
x J

_
d[f] e

i
2
R
d
d
xf(x)f(x)
, (4.188)
and it is evident that
f)
J

=
_
d[f] f e

i
2
R
d
d
x f(x)f(x)
= 0 )
J

= . (4.189)
Hence we recover the relation
W[J]
J(x)

J=J

= , (4.190)
and also we have
e
i[]
= e
iS[]
_
d[f] e

i
2
R
d
d
x f(x)f(x)
. (4.191)
The Gaussian functional integral is evaluated as usual giving
_
d[f] e

i
2
R
d
d
x f(x)f(x)
= (det )

1
2
(det
0
)
1
2
, (4.192)
where the normalisation is chosen so that when V = 0 the integral will be one, which requires

0
=
2
. (4.193)
We may now take the logarithm to obtain
[] :=
_
d
d
x V
e.
() =
_
d
d
x V ()
i
2
(log det log det
0
.) (4.194)
The additional term involving det / det
0
is then the rst quantum correction to the classical
potential.
To calculate the determinants we use the relation log det = tr log . The trace is obtained
by summing over the eigenvalues of the operator . Since V

() is a constant the eigenfunctions

are given by e
ikx
, with corresponding eigenvalues k
2
+ V

(). If the system is in a large box

of size L then, with suitable boundary conditions, k = O(L
1
) and is discrete. As in statistical
mechanics as L we may replace the sum over k by

k
1
_
d
d
k
(2)
d
, (4.195)
for 1 = O(L
d
) the volume of the box. This gives the result, factoring o 1,
V
e.
() = V ()
i
2
_
d
d
k
(2)
d
_
log(k
2
+V

() i) log(k
2
i)
_
, (4.196)
where the singularities are treated just as in Feynman propagators. Performing a Wick
rotation by analytically continuing the contour of k
0
:
V
e.
() = V () +
1
2
_
d
d
k
(2)
d
_
log(k
2
+V

()) log k
2
_
, (4.197)
where now k
2
> 0.
To carry out the integral, we use an integral representation for a, b > 0:
log
a
b
=

_
0
d

(e
a
e
b
) , (4.198)
and then, for a > 0,
_
d
d
k
(2)
d
e
ak
2
=
1
(4a)
d
2
. (4.199)
70
Hence we may compute
V
e.
() = V ()
1
2
_
d
d
k
(2)
d

_
0
d

_
e
(k
2
+V

())
e
k
2
_
= V ()
1
2
1
(4)
d
2

_
0
d

d
2
1
_
e
V

()
1
_
. (4.200)
This is convergent for 0 < d < 2. With an integration by parts and the standard Gamma
function integral
V
e.
() = V () +
1
d
1
(4)
d
2

_
0
d
_
d
d

d
2
_
_
e
V

()
1
_
= V () +
1
d
1
(4)
d
2
V

()

_
0
d

d
2
e
V

()
= V () +
1
d
(1
d
2
)
(4)
d
2
V

()
d
2
= V ()
(
d
2
)
2(4)
d
2
V

()
d
2
. (4.201)
Note that this makes sense only when V

() > 0.
The result is valid for general d but (
d
2
) has poles at d = 0, 2 and d = 4 reecting
divergencies. In particular

d
2
_

1
2

_
1 +
3
4

_
, d = 4 . (4.202)
This pole is removed by adding the counterterm potential
V
c.t.
() =

32
2

()
2
. (4.203)
where the arbitrary mass has been introduced to ensure dimensional consistency since V ()
has dimension d and V

() has dimension 2. Since V

()
d
2
V

()
2
_
1
1
2
log V

()
_
then
V
e.,ren.
() =
_
V
e.
() +V
c.t.
()
_

0
= V () +
1
64
2
V

()
2
_
log
V

()

2

3
2
_
,
2
= 4 e

2
. (4.204)
It is important to recognise that V
e.,ren.
() is, as a consequence of the necessity of adding
counterterms to ensure niteness, arbitrary up to the addition of a nite quartic polynomial in
. Thus in this result the coecient of V

()
2
, as opposed to that for V

()
2
lnV

(), has no
physical signicance. This freedom can be removed by specifying the derivatives V
(n)
e.,ren.
(0), n =
1, 2, 3, 4, as the parameters on which the theory depends (if the theory has a symmetry,
only n = 2, 4 are relevant). Note that, from the relation to [], we have V
(n)
e.,ren.
(0) =

n
(0, . . . , 0)
ren.
.
It is the eective potential whose minima determine the ground state for a quantum eld
theory. These may be dierent from those for the classical theory, even if the parameters
describing each are matched in some way. The nite V
e.,ren.
() depends on the arbitrary scale
which is introduced through regularisation of the Feynman integrals for vacuum diagrams
that are used to calculate it. In consequence it satises an RG equation which has the typical
form, if there is one coupling g,
_

+(g)

g

(g)

_
V
e.,ren.
() = 0 . (4.205)
71
5 Gauge Theories
Up to now we have mainly used scalar eld theory as an example of a quantum eld theory;
the point is that realistic theories involve gauge elds and fermions. Associated with each
gauge eld is a gauge group, which is usually a continuous Lie
37
group. In the most important
examples
38
, these are
Theory: QED Weinberg-Salam model QCD,
Gauge Group: U(1) SU(2) U(1) SU(3).
The gauge elds are vector elds, so they have a Lorentz index A

(x), and they belong to

the Lie algebra g of the Lie group G. Denoting the set of all vector elds taking values in g
by /, we can write this as
A

/. (5.1)
In a formal sense, the gauge group ( is dened by
(

x
G
x
, (5.2)
i.e. an element of ( is a map from spacetime points to elements of the Lie group G (this
becomes precise when spacetime is approximated by a lattice).
5.1 Brief Summary of Lie Algebras and Gauge Fields
For a continuous Lie group G the elements g() depend on parameters
r
, r = 1, . . . , dimG,
where everything is continuous and dierentiable with respect to . (When considering gauge
groups, the parameters are taken to be functions of the spacetime points x.) The group
structure of G means that for any given ,

,
g()g(

) = g(

) , (5.3)
for some

. Generally g(0) = e, the identity. A Lie algebra g is a vector space equipped with
a Lie bracket [, ], so that
X, Y g [X, Y ] g . (5.4)
The Lie bracket is antisymmetric and satises a Jacobi identity.
[X, Y ] = [Y, X] , [[X, Y ], Z] + [[Z, X], Y ] + [[Y, Z], X] = 0 . (5.5)
We can conjugate elements of the Lie algebra by group elements,
g
1
Xg g , g G. (5.6)
For a small change in g
g
1
dg g , (5.7)
and there is also an exponentiation map from the Lie algebra to a one parameter subgroup of
Lie group:
exp : g G; X g
t
= exp(tX) . (5.8)
We can take a basis t
a
, a = 1, . . . , dimG, for g, so that any X g can be written as
X = X
a
t
a
, and dene structure constants f
abc
by
[t
a
, t
b
] = f
abc
t
c
, (5.9)
where the basis may be chosen so that f
abc
is totally antisymmetric. With a basis we have
g()
1
dg() = t
a
l
ar
()d
r
, (5.10)
37
Lie, Sophus (1842-1899)
38
Weinberg, Steven (1933-), Salam, Abdus (1926-1996), Nobel Prizes 1979
72
and we may choose l
ar
(0) =
ar
. It is easy to see that
g
0
g() = g(

) l
ar
()d
r
= l
ar
(

)d

r
, (5.11)
for xed g
0
. Under the change of variables

, the Jacobian det[

/] = det l()/ det l(

)
and we may dene invariant integration over the group by
d(g) = det l()

r
d
r

_
d(g) f(g) =
_
d(g) f(g
0
g) . (5.12)
For compact groups
_
G
d(g) is nite.
With a basis then for any X g there is a corresponding matrix dened by
[t
a
, X] = T
ab
(X)t
b
. (5.13)
Clearly
t
ab
(t
c
) = f
acb
. (5.14)
The matrices T(X) = T
ab
(X) form a representation of the Lie algebra since, as a consequence
of the Jacobi identity, [T(X), T(Y )] = T([X, Y ]), where we use standard matrix multiplica-
tion. This is called the adjoint representation of the Lie algebra, the matrices have dimension
dimGdimG where the parameters have dimension dimG (note that dimSU(2) = 3, whereas
dimSU(3) = 8). The adjoint representation can be extended to G by g
1
t
a
g = D
ab
(g)t
b
.
There is also an quadratic form X Y which is invariant under a conjugation
X Y = g
1
Xg g
1
Y g . (5.15)
An equivalent relation for the Lie algebra, obtained for g g
t
= exp(tZ) and letting t 0, is
[X, Z] Y +X [Y, Z] = 0 . (5.16)
If G is compact the scalar product is positive denite. By a suitable change of basis one can
then arrange
t
a
t
b
=
ab
. (5.17)
For simple Lie groups, those not of the form GG

, the scalar product is unique.

A gauge transformation of a gauge eld A

(x) by an element of the gauge group g(x), i.e.

g(x) G for each x, is dened by
A

(x) A
g

(x) := g(x)
1
A

(x)g(x) +g(x)
1

g(x) . (5.18)
In the case of G = U(1), we can write g(x) = e
i(x)
, so that
A
g

(x) = A

(x) +i

(x) . (5.19)
The gauge eld may be expanded
A

= A
,a
t
a
, (5.20)
for a = 1, . . . , dimG, and plays the role of a connection similar to general relativity; the
corresponding curvature is
f

+ [A

, A

] . (5.21)
Under a gauge transformation,
f

(x) F
g

(x) = g(x)
1
F

(x)g(x) , (5.22)
where F
g

(x) is the curvature formed from the transformed connection A

g

(x). Unlike A

, F

transforms homogeneously under gauge transformations.

With the connection we can dene a covariant derivative acting on elds belonging to a
representation space V

for the Lie algebra

d

+A

, (5.23)
73
where
A

= A
,a
T
a
, (5.24)
forming T
a
form a representation of the Lie algebra, i.e. they are matrices acting on V

satisfying [T
a
, T
b
] = f
abc
T
c
. Under a gauge transformation
g
= R(g), where R(g) is the
matrix corresponding to g in the representation dened on V

. The gauge transformations are

dened so that (D

)
g
= D
g

g
. It is straightforward to see that we can write
[D

, D

] = F
,a
T
a
. (5.25)
The gauge elds belong to the representation space for the adjoint representation so that the
covariant derivative becomes
(D

)
a
=

F
,a
+ [A

, F

]
a
=

F
,a
+T
ab
(A

)F
,b
. (5.26)
Also for an innitesimal gauge transformation
A

= D

, (5.27)
for D

the adjoint covariant derivative and belongs to the Lie algebra for (, formed by elds

a
(x) with
a
(x)t
a
g.
The curvature satises the crucial Bianchi
39
identity, which may be obtained by using the
Jacobi identity for [D

, [D

, D

]],
d

+D

+D

= 0 , (5.28)
for D

the adjoint covariant derivative as above.

5.2 Gauge Fields in Quantum Field Theory
The simplest gauge invariant scalar which forms a Lagrangian is
L =
1
4g
2
F

, (5.29)
where g is introduced as a overall coecient and is the coupling for the gauge theory. The
corresponding action is then
S[A] =
_
d
d
x L(x) . (5.30)
Clearly S[A] = S[A
g
]. There is no constraint on the dimension. The variational equations
S = 0 give the equations of motion
D

= 0 . (5.31)
These equations do not depend on g which is irrelevant classically.
Let us suppose that A

(x) is a solution to these equations, then, as a consequence of

gauge invariance, A
g

(x) is also a solution for any element g(x) of the gauge group. Since
g(x) is unconstrained the dynamical equations do not have unique solutions given some initial
conditions on A

(x) and its time derivative at some initial time. It follows that gauge theories
have redundant degrees of freedom which do not have any dynamical role. At the classical level
this is not too important (for abelian electrodynamics F

is gauge invariant), but it leads to

potential problems in quantisation. The dynamical variables in this gauge theory really belong
to the equivalence class of gauge elds modulo gauge transformations
//( = A

A
g

: A

/, g ( . (5.32)
The set A
g
for all g ( forms the orbit of A under the action of (. The space of gauge elds
/ is a linear space, if A

, A

/ then so is aA

+ a

, and so is topologically trivial but

the space of orbits or //( is nontrivial in a topological sense. The dynamical variables in a
39
Bianchi, Luigi (1856-1928)
74
gauge theory are coordinates on //( so that there is a well dened initial value problem and
quantisation can be achieved.
To exhibit the potential diculties in quantising we rst consider a canonical approach.
When quantising a classical theory given by a Lagrangian L(q
i
, q
i
), one rst denes the
conjugate momenta by
p
i
=
L
q
i
, (5.33)
which gives p
i
(q, q). It is crucial that this should be invertible so that we can write q
i
(q, p). For
a gauge theory we let q
i
A

(x). Then the conjugate momenta are given by

L

A
i
=
1
g
2
F
0i
,
L

A
0
= 0 . (5.34)
L does not depend on

A
0
, but if we vary A
0
, we get from S = 0,
d
i
F
0i
= 0 , (5.35)
which is a non dynamical equation, without t derivatives, similar to



E = 0 in vacuum
electrodynamics. There is no momentum conjugate to A
0
, it is not a dynamical variable, but i
acts as a Lagrange multiplier enforcing a constraint.
Another related problem arises when expanding S[A], which contains quadratic, cubic and
quartic terms in A. Consider just the quadratic terms then
S[A] =
1
2g
2
_
d
d
x
_
A

(A)

+O(A
3
)
_
, (5.36)
where
(A)

=
2
A

. (5.37)
In quantum eld theory then, for a perturbation approach in terms of Feynman diagrams, we
need to invert to dene the Feynman propagator. But it is easy to see that for any function

()

= 0, (5.38)
which is a consequence of S[A
g
] = S[A]. We cannot then invert in a straightforward fashion
so that there is no direct perturbative expansion with the cubic and higher order terms dening
the interaction.
These issues no longer arise if the dynamical variables are regarded as belonging to //(
instead of /. In a functional integral approach this requires that the quantum eld theory is
dened by
_
A/G
d[A] e
iS[A]
, (5.39)
where is some measure on //( that we need to dene. In general we require that
_
A
d[A] =
_
G
d[g]
_
A/G
d[A] , (5.40)
where the integration over the gauge group ( is required to satisfy
_
G
d[g] F[g] =
_
G
d[g] F[g
0
g] for any g
0
( . (5.41)
There is a standard prescription for constructing the required measure on //( which is to
consider an integration over all elds d[A] and impose a gauge xing condition F(A) = 0, where
F(A) depends locally on A

(x) and its derivatives and for each x F(A) g. We assume that
F(A
g
) = 0 , for all A / for some unique g ( . (5.42)
With these requirements the gauge condition F(A) = 0 selects a unique gauge eld A

on each
gauge orbit. Actually this is not possible in general for any A, because //( is topologically
75
non trivial, but it is feasible for small A, and this is sucient in perturbation theory. It is also
essential for the functional integral to be just on //( that the results are independent of the
particular choice of the gauge xing function F(A).
To achieve these requirements we suppose that
_
A/G
d[A] =
_
A
d[A] [F(A)] M[A] , (5.43)
where we have included a functional delta function imposing the gauge xing condition and
also have introduced the function M[A], dened for A satisfying F(A) = 0, which compensates
for the choice of F. This is achieved by requiring
_
G
d[g] [F(A
g
)] M[A] = 1 (5.44)
for any given A. Using invariance of the group integration M(A
g
0
) = M[A]. By assumption
(5.42) there exists a unique g
0
such that F(A
g
0
) = 0 so that the only contribution to the integral
is for g g
0
. It is easy to see by letting A A
g
0
it is sucient to assume g
0
= e, the identity.
We then note that, since then A
g

= A

+D

where D

is the adjoint covariant derivative,

F[A] = 0 F[A
g
] = F

(A)

=:
gh
, (5.45)
where we dene

gh
= F

(A)

. (5.46)
In the neighbourhood of the identity, with a suitable normalisation d[g]

ge
= d[], where d[]
denotes the functional integration measure over elds (x) g which is a linear space. Hence
we have
_
G
d[g] [F(A
g
)] =
_
d[] [
gh
] = (det
gh
)
1
, (5.47)
so that the integration has been reduced to one over elds belonging to the Lie algebra g of
the gauge group G.
gh
is called the ghost operator, for reasons which will become apparent
later on and in (5.44) we must take
M[A] = det
gh
. (5.48)
It is important to note that if
gh
= 0 for some then this is a signal that the gauge
xing condition does not x the gauge uniquely as in (5.42).
The functional integral over the gauge elds may now be written as
_
A
d[A] e
iS[A]
=
_
A
d[A]
_
G
d[g] [F(A
g
)]M[A] e
iS[A]
=
_
G
d[g]
_
A
d[A] [F(A
g
)]M[A] e
iS[A]
=
_
G
d[g]
_
A
d[A
g
] [F(A
g
)]M[A
g
] e
iS[A
g
]
=
_
G
d[g]
_
A
d[A] [F(A)]M[A] e
iS[A]
, (5.49)
where we use that d[A], M[A], S[A] are invariant under A A
g
. Hence nally we may identify
_
A/G
d[A] e
iS[A]
=
_
A
d[A] [F(A)] det
gh
e
iS[A]
, (5.50)
with
gh
determined by the choice of gauge xing function F(A).
5.3 Example: Ordinary Finite Integral
To illustrate how this works we discuss an example for an ordinary nite integral which realises
the same issues as for a functional integral over gauge elds. For a vector x R
n
we consider
the following integral:
_
d
n
x f(x) . (5.51)
76
We assume f(x) is invariant under the symmetry group SO(n), so that
f(x) = f(x

) for all x

= Rx, R SO(n) . (5.52)

Since for
x x

= Rx, d
n
x

= d
n
x, (5.53)
as det R = 1 the integral is invariant. SO(n) here plays the role of a gauge group and the orbit
of x under the action of SO(n) is a sphere of radius [x[.
For this example we know how to evaluate the integral since
f(x) =

f(r) , r = [x[ , (5.54)
we may integrate over angles giving
_
d
n
x f(x) = S
n

_
0
dr r
n1

f(r) , (5.55)
where S
n
=
2

n
(
n
2
)
is the area of an (n 1)-dimensional unit sphere.
We now approach this integral in the same fashion as the gauge eld functional integral
by using a gauge xing condition. In this case it is clear that by a rotation R we can always
arrange
x x
0
:= r(0, 0, . . . , 0, 1) . (5.56)
This gauge xing condition is then to set x = x
0
which can be realised by introducing

_
F(x)
_
:= (x
n
)
n1

i=1
(x
i
) , (5.57)
into the integral. To ensure complete xing of the gauge freedom for x we also require x
n
> 0.
Just as before, we also introduce a function M(x) by the requirement
_
SO(n)
d(R) (F(Rx)) M(x) = 1 , (5.58)
where d(R) is the invariant measure for integration over SO(n). In this example the gauge
xing condition does not determine R uniquely since there is a subgroup SO(n 1) SO(n),
acting on the rst (n 1)-components, such that for R

SO(n 1) then R

x
0
= x
0
. For
the Lie algebra then if t
a
a = 1, 2, . . . ,
1
2
n(n 1), are antisymmetric matrices forming the
generators of SO(n) then we may decompose t
a
=

t
i
, t
s
so that

t
i
x
0
= 0 , i = 1, . . . ,
1
2
(n 1)(n 2) , (5.59)
with

t
i
the generators for the SO(n 1) subgroup, while the remaining generators t
s
then
satisfy
t
s
x
0
= r(0, . . . , 1
..
s

th place
, . . . , 0) , s = 1, . . . , n 1 . (5.60)
To calculate M(x) we need only consider rotations R of the form
R = (1 +
s
t
s
)R

neglecting O(
s

s
) . (5.61)
For such R
Rx
0
= r(
1
, . . . ,
n1
, 1) +O(
s

s
) , (5.62)
and therefore

_
F(Rx
0
)
_
=
n1

i=1
(r
i
) . (5.63)
77
The integration measure may be dened so that
d(R) = d(R

)
n1

s=1
d
s
_
1 +O(
s

_
. (5.64)
The required integral then becomes
_
SO(n)
d(R)
_
F(Rx
0
)
_
=
_
SO(n1)
d(R

)
_
n1

s=1
d
s
(r
s
) =
V
SO(n1)
r
n1
, (5.65)
where V
SO(n1)
=
_
d(R

) is the volume of the group SO(n1). Hence we choose the function

M to be
m(x) =
r
n1
V
SO(n1)
. (5.66)
Note that M(Rx) = M(x).
Following the same procedure as used for the gauge invariant case the SO(n) invariant
integral can be rewritten as
_
d
n
x f(x) =
_
d
n
x
_
SO(n)
d(R)
_
F(Rx)
_
M(x) f(x)
=
1
V
SO(n1)
_
SO(n)
d(R)
_
d
n
x
_
F(Rx)
_
r
n1
f(x
0
)
=
1
V
SO(n1)
_
SO(n)
d(R)
_
d
n
x (x
n
)

i
(x
i
) r
n1
f(x
0
)
=
V
SO(n)
V
SO(n1)

_
0
dr r
n1
f(x
0
) , (5.67)
where the nal integration is just over x
n
= r > 0. This is equal to the result obtained by
integrating over angles if we use
S
n
=
V
SO(n)
V
SO(n1)
. (5.68)
5.4 Introduction of Ghost Fields
Returning to the functional integral of a gauge eld theory we consider then
_
A
d[A] [F(A)] det
gh
(A) e
iS[A]
, (5.69)
where
S[A] =
1
4g
2
_
d
d
x F

(x) F

(x) , (5.70)
and we note that in general the ghost operator depends on the gauge eld A. If it does not
the corresponding determinant can be factored out of the functional integral and absorbed in
the overall arbitrary normalisation constant. It remains to show how the expansion of this
functional integral can be expressed in term of appropriate Feynman rules. To achieve this we
seek a representation in which all factors are present in an exponential
First of all, rewrite
[F(A)] =
_
d[b] e
i
g
2
R
d
d
x b(x)F(A)
, (5.71)
where b(x) is a bosonic eld, which like F

and F(A), belongs to the Lie algebra. This

functional integral is the functional analogue of the usual formula for the -function. The
prefactor of
1
g
2
has been introduced for later convenience. Also let us rewrite the determinant
as a function integral over Grassmann elds, as shown section 2.3,
det
gh
(A) =
_
d[ c]d[c] e
iS
gh
[ c,c,A]
, (5.72)
78
where
S
gh
[ c, c, A] =
1
g
2
_
d
d
x c(x)
gh
(A)c(x) . (5.73)
c and c are here Fermion scalar elds, which also belong to the Lie algebra, and are treated as
independent elds. They are called ghost elds since they are unphysical because they do not
correspond to physical particles, having the wrong statistics.
The function integral then becomes
_
d[A]d[b]d[ c]d[c] e
iS
q
[A,b,c, c]
, (5.74)
where the quantum action is given by
S
q
[A, b, c, c] =
1
g
2
_
d
d
x
_

1
4
F

(x) F

(x) +b(x) F(A) + c(x)

gh
(A)c(x)
_
. (5.75)
The additional auxiliary eld b adds an additional degree of freedom while the Grassmann
ghost elds c, c subtract two degrees of freedom so that the net degrees of freedom correspond
to that for gauge elds modulo gauge transformations.
There is a further generalisation which extends the notion of gauge xing introduced above.
It is sucient, in order to produce a well dened theory, to have a family of gauge xing
conditions which are then summed or integrated over. Thus we consider a gauge xing condition
F(A) = f where f(x) L
G
and we integrate with Gaussian weight function
[F(A)]
_
d[f] [F(A) f] e

i
g
2
1
2
R
d
d
x f(x)f(x)
= e

i
g
2
1
2
R
d
d
x F(A)F(A)
=
_
d[b] e
i
g
2
R
d
d
x(b(x)F(A)+
1
2
b(x)b(x))
, (5.76)
where in the last line the same expression is obtained by extending the b-functional integral by
a quadratic term. This is easily evaluated, up to an irrelevant overall constant, by completing
the square, assuming a convenient choice of scale for the b functional integral. The choice of
gauge now depends on F(A) and . To go back to the previous case when F(A) = 0 requires
taking the limit 0.
In consequence the quantum action has been extended to
S
q
[A, b, c, c] =
1
g
2
_
d
d
x
_

1
4
F

(x) F

(x) +b(x) F(A) +

1
2
b(x) b(x) + c(x)
gh
(A)c(x)
_
.
(5.77)
For many calculations it is sucient to consider a very simple choice, the linear covariant
gauge
f(A) =

. (5.78)
This maintains Lorentz invariance. In this case F

(A)

and

gh
=

. (5.79)
5.5 Feynman Rules for Gauge Theories
Let us try and set up a perturbative expansion; although it can be considered as an independent
eld it is simplest to eliminate b essentially by completing the square. We can then write the
action as
S
q
[A, c, c] =
1
g
2
_
d
d
x
_

1
4
F

1
2

c D

c
_
. (5.80)
This allows the Feynman rules to be derived in an essentially straightforward fashion, with
diagrams involving lines and vertices for the elds A

, c and c.
79
Look rst of all at the quadratic (free) part of the action:
S
q
[A, c, c]
quad
=
1
g
2
_
d
d
x
_

1
2

)
1
2

c
_
=
1
g
2
_
d
d
x
_
1
2
A

_
1
1

_
A

+ c
2
c
_
=
_
d
d
x
_
1
2
A

c
2
c
_
, (5.81)
where

=
2

+
_
1
1

, (5.82)
and in the last step we removed the overall factor 1/g
2
by redening
A

gA

, c, c gc, g c . (5.83)
From this we may obtain the two-point functions
A
a
(x) A
b
(y)), c
a
(x) c
b
(y)) . (5.84)
As was shown in the case of scalar elds the propagators are just the Green functions for the
dierential operators that appear in the quadratic part of the action except there are now
Lorentz and group indices to take account of
i

x
A
a
(x) A
b
(y)) =
ab

d
(x y) , (5.85)
i
2
x
c
a
(x) c
b
(y)) =
ab

d
(x y) . (5.86)
The standard procedure to solve these equations is by considering Fourier transformations:
_
d
d
x e
ipx
A
a
(x) A
b
(0)) =: i
ab

F,
(p) , (5.87)
_
d
d
x e
ipx
c
a
(x) c
b
(0)) =: i
ab

F
(p) , (5.88)
so that we have to solve

_
p
2

_
1
1

_
p

F,
(p) =

, (5.89a)
p
2

F
(p) = 1 . (5.89b)
These are pretty straightforward to invert:

F,
(p) =

p
2
i
+ (1 )
p

(p
2
i)
2
,

F
(p) =
1
p
2
i
.
We therefore get the following propagators:
For a gauge eld propagator, introduce a wavy line:

p
, a , b
It corresponds to
i
ab
_

p
2
i
+ (1 )
p

(p
2
i)
2
_
. (5.90)
We can see that the propagator gets simplied by choosing = 1 (the Feynman gauge).
However, one can leave the parameter free in any calculation to check if the nal results
for physical quantities are independent of .
80
For a ghost propagator, introduce a dashed line:

p
a
c
b
c
which corresponds to
i
ab
1
p
2
i
. (5.91)
Note that
p

F,
(p) = p

F
(p) . (5.92)
In order to get the Feynman rules for vertices, we require cubic and higher order terms in
S
q
, which, after rescaling of the elds, are given by
S
q
[A, c, c]
int
=
_
d
d
x
_
gf
abc
A

a
A

A
c
g
2
1
4
f
abe
f
cde
A

a
A

b
A
c
A
d
gf
abc

c
a
A

b
c
c
_
,
(5.93)
where we used [X, Y ]
a
= f
abc
X
b
Y
c
and D

c =

c + [A

, c]. For an interaction involving

a derivative of a eld, such as

(x), then in the Feynman rules for the associated vertex

ip

where p

is the ingoing momentum to a vertex at x.

40
The rst term in the interaction
part of the action describes an interaction of the form
, b
q
, a
@R
p
, c
r
We have a cyclic symmetry (r, , c p, , a q, , b), and the vertex is given by
Gf
abc
(r

+p

+q

) , (5.97)
where we have 6 = 3! terms because of the cyclic symmetry. The second term in the interaction
part gives
, b
, a
@R
, c
@I
, d

Note that the interaction term is symmetric under the exchange A

a
A

b
and A
c
A
d
,
which cancels the factor of four. The vertex is
ig
2
(f
abe
f
cde
(

) +f
cae
f
bde
(

) +f
bce
f
ade
(

)) .
(5.98)
We also have an interaction involving ghost elds and a gauge eld:
40
Supposing we had
L
I
=
g
6

3
, (5.94)
the interaction vertex would be

R
@
q
p
q
ig, p + q + r = 0
Supposing we had the following

g
2

, (5.95)
in this case we would get for such a vertex
i(ig)(p + q + r)

= 0 , (5.96)
because the interaction is a total derivative.
81
, b
q
c, a - p
c, d
r
ig(ip

)f
abd
. (5.99)
It is important to note that the sign of the c c propagator and the Ac c vertex is not signicant
but their relative sign is.
In the Feynman rules cubic vertices are then proportional to a momentum, while the quartic
vertices are independent of any momenta. Taking this into account the calculation of the overall
degree of divergence D for a Feynman integral proceeds in essentially the same fashion as that
followed before for a scalar theory. The formula is
d = 4 E
A
E
gh
. (5.100)
Note that in the Feynman rules there is a minus sign for ghost loops, which is because they
are fermionic, anticommuting elds.
Now one can in principle use these rules to calculate Feynman amplitudes, but the indices
make life more dicult; calculations become more dicult practically.
5.6 Canonical Approach
Although the functional integral is ne for deriving Feynman rules for gauge theories it is
necessary, in order to construct operators and the associated Hilbert space, to consider the
canonical approach to quantisation involving dynamical variables and their associated conjugate
momenta which are then promoted to operators. As a digression we outline how this proceeds.
With the linear covariant gauge the full quantum Lagrangian density can be expressed just
in terms of the elds and their rst derivatives in the form
L
q
=
1
4
F

b A

+
1
2
b b

c D

c . (5.101)
In this case if we were to consider this as a starting point for canonical approach to quantisation,
we would nd
L
q

A
i
= F
0i
,
L
q

b
= A
0
,
L
q
c
=
0
c =

c ,
L
q

c
= D
0
c . (5.102)
Hence A
i
, F
0i
and b, A
0
and c,
0
c form conjugate pairs like q
i
, p
i
so that we may impose on
the associated operators canonical commutation, or anti-commutation, relations at equal times,
_

A
i
(x),

F
0i
(0)]
e.t.
= i
d1
(x) ,
_

b(x),

A
0
(0)]
e.t.
= i
d1
(x) ,
_
D
0
c(x),

c(0)
_
e.t.
= i
d1
(x) ,
_

c(x), c(0)
_
e.t.
= i
d1
(x) . (5.103)
For simplicity if we consider the free theory, when F

, D

c, which
is a starting point for a perturbative treatment. The associated Fock space of states is also
the space on which operator elds are initially dened. In this case the equations of motion
require

b = 0 ,

+ b = 0 ,
2
c = 0 ,
2
c = 0 . (5.104)
These also require

2
A

= (1 )

b ,
2
b = 0 . (5.105)
The operator elds then satisfy
_

(x),

A

(0)

= i

D(x)

1 +i(1 )

E(x)

1,
_

(x),

b(0)

= i

D(x)

1,
_

b(x),

b(0)

= 0 ,
_
c(x),

c(0)
_
= iD(x)

1. (5.106)
Here
D(x) =
i
(2)
3
_
d
d1
k
2[

k[
_
e
ikx
e
ikx
_

k
0
=|

k|
,
2
E(x) = D(x) . (5.107)
Since
2
D(x) = 0 the commutation relations are compatible with the free equations of motion
and also, since
t
D(x)[
t=0
=
d1
(x), with the equal time commutation relations.
82
5.7 BRS Symmetry
The quantisation process has introduced additional elds into the theory, so that the space
of states necessary after quantisation is now larger than would correspond to the expected
physical degrees of freedom. The larger space contains negative norm states and ghost states
which violate the spin-statistics theorem. It is vital that we can identify what are the physical
states within this space and that these form a Hilbert space with positive norm so that
ordinary quantum mechanics is valid.
This is possible because of an additional symmetry, the so-called BRS symmetry
41
, which
is related to gauge invariance. BRS symmetry is an extension of the gauge symmetry for the
classical action to the quantum action including additional elds. The Lagrange density,
choosing a linear gauge xing F(A) = F

for simplicity becomes

L
q
=
1
4
F

(x) F

(x) +b(x) F

(x) +
1
2
b(x) b(x) + c(x) F

c(x) . (5.108)
Clearly only the rst term is gauge invariant, where for an innitesimal gauge transformation
A

= D

= [F

, ] (F

) = 0 . (5.109)
Although the additional terms present in L
q
are not gauge invariant, there is a residual sym-
metry: L
q
is invariant under BRS symmetry. To exhibit this we dene an operator s acting on
the elds A

, c, c and b, that relates bosons and fermions

42
:
sA

= D

c =

c + [A

, c] ,
sc =
1
2
[c, c] ,
where [c, c]
a
= f
abc
c
b
c
c
is non-zero because c
b
c
c
= c
c
c
b
. What is non trivial is that s is
nilpotent s
2
= 0. The action of s anti-commutes with Fermion elds so that
s(cX) = (sc)X csX , (5.111)
and similarly for c. To prove s
2
= 0 we verify rst
s
2
A

= D

1
2
[c, c] + [D

c, c] = 0 , (5.112)
as a consequence of
D

[X, Y ] = [D

X, Y ] + [X, D

Y ] , (5.113)
and furthermore
[c, D

c] = [D

c, c] , (5.114)
since [c, D

c]
a
= f
abc
c
b
(D

c)
c
= f
acb
(D

c)
c
c
b
= [D

c, c]
a
for anticommuting elds (whereas for
bosonic elds [X, Y ] = [Y, X], of course). Secondly
s
2
c =
1
4
[[c, c], c]
1
4
[c, [c, c]] =
1
2
[c, [c, c]] = 0 , (5.115)
because
[c, [c, c]]
e
= f
cde
c
c
[c, c]
d
= f
cde
f
abd
c
c
c
a
c
b
= f
cde
f
abd
c
[c
c
a
c
b]
= 0 , (5.116)
by the Jacobi identity
f
de[c
f
ab]d
= 0 . (5.117)
We can extend s to the other elds by dening
s c = b , sb = 0 , (5.118)
41
Named after Carlo M. Becchi, Alain Rouet, Raymond Stora, who came up with it; sometimes referred to as
BRST symmetry because of a Russian, Igor Viktorovich Tyutin (1940-), who was supposed to have also developed
the concept, but his paper was never published, an English version has now appeared in arXiv:0812.0580.
42
So it is reminiscent of supersymmetry, which was discovered slightly earlier, although that relates physical
particles.
83
and acting on these elds it is trivial to see that s
2
= 0. Thus we have dened the action of s
on the basic elds (A

, c, c, b) so that in general
s
2
= 0 . (5.119)
With this choice the quantum Lagrange density in linear gauges can then be written in
the form
L
q
=
1
4
F

(x) F

(x) s (x) , (x) = c(x) F

(x) +
1
2
c(x) b(x) , (5.120)
as is easy to see:
s (x) = b(x) F

(x) c(x) F

c(x)
1
2
b(x) b(x) , (5.121)
using that s anticommutes with c, c. The rst classical term in L
q
is gauge invariant since
sA

= D

c is just a innitesimal gauge transformation, albeit with c a Grassmann element of

the Lie algebra. Since s
2
= 0 we have the important property
sL
q
= 0 . (5.122)
In mathematics with an operation s acting on some vector space, or ring (where elements
can be multiplied as well as added), such that s
2
= 0 then one sets up the cohomology of s
on this space by considering all elements X such that sX = 0. These elements are called
closed. If X = sY then X is trivially closed and elements of this form are exact. There is an
equivalence relation for closed elements such that X X

if XX

is exact, i.e. XX

= sY
for some Y . The cohomology class dened by s is dened by X : sX = 0, X X+sY . Thus
L
q
is exact under the BRS transformation s and belongs to the same equivalence class as the
classical L.
The BRS transformation can be used to construct a symmetry under which L is invariant
by dening

_
A

, c, b, c
_
= s
_
A

, c, b, c
_
, (5.123)
where is a Grassmann parameter. It is easy to see that

L
q
= 0 and also

2
= 0. Another
relevant symmetry of L
q
is obtained by the transformations

c = c ,

c = c , (5.124)
since c occurs only in association with c and where we take into account that c, c are regarded
as real elds. These transformations on the ghost elds generate a one dimensional symmetry
group which leads to conservation of ghost number.
We apply Noethers theorem for the BRS transformations

and also for the the

transformations on c, c. For simplicity we assume here the linear covariant gauge and write L
q
in a form where there are only rst derivatives of the elds
L
q
=
1
4
F

b A

c D

c +
1
2
b b . (5.125)
For (x) depending on x the terms involving derivatives come from

c =

1
2
[c, c] ,

c =

b , (5.126)
and

= D

c +[F

, c] . (5.127)
In calculating

L
q
the terms without derivatives sum up to zero so the relevant answer
becomes

L
q
=

_
F

c +b D

c
1
2
[c, c]
_
. (5.128)
This gives the BRS current
j

BRS
= F

c b D

c +

c
1
2
[c, c] . (5.129)
84
In a similar fashion letting (x) depend on x

L
q
=

c c c D

c
_
, (5.130)
so the corresponding ghost current is
j

gh
=

c c c D

c . (5.131)
These results lead to a conserved quantities, the BRS charge and the ghost charge
Q
BRS
=
_
d
d1
x j
0
BRS
(x) , Q
gh
=
_
d
d1
x j
0
gh
(x) . (5.132)
In the quantum eld theory the elds become operators acting on a Hilbert space H
which can be generated by the action of the elds on the vacuum [0). We require that

A
a

A
a
, c
a

= c
a
,

c
a

c
a
(note that with A
a
= A
a

, c
a
= c
a

, c
a
= c
a

, b
a
= b
a

then
L
q
= L
q

) and then there are Hermitian BRS and ghost charges

Q
BRS
=

Q
BRS

,

Q
gh
=

Q
gh

. (5.133)
We must then have
_

Q
gh
, c

= i c ,
_

Q
gh
,

= i

c , (5.134)
and, for anti-commuting,
_

Q
BRS
, X

= i

X . (5.135)
It is crucial that the BRS charge satises

Q
BRS
2
= 0 , (5.136)
corresponding to s
2
= 0 and, since

Q
BRS
increases the ghost charge by one,
_

Q
gh
,

Q
BRS

= i

Q
BRS
. (5.137)
Since these are conserved we must have
_

H,

Q
BRS

= 0 ,
_

H,

Q
gh

= 0 . (5.138)
The aim now is to dene the physical space of states. The initial Hilbert space H has
negative and zero norm states, otherwise nilpotent operators like

Q
BRS
would be trivial. If you
quantise a gauge theory in a Lorentz invariant fashion negative norm states are essentially
inevitable. The essential assumption is that physical states correspond to the cohomology
classes dened by

Q
BRS
.
First dene the subspace H
0
H by

Q
BRS
[) = 0 [) H
0
. (5.139)
Clearly this is a subspace because H
0
is the kernel of

Q
BRS
. Since

Q
BRS
is conserved the
space H
0
is invariant under time evolution. The essential assumption is that H
0
is positive
semi-denite or
[) 0 [) H
0
. (5.140)
For any positive semi-denite space a proper Hilbert space can be dened by considering
equivalence classes
[) [

) if [

) [) = [) , [) = 0 . (5.141)
By this construction if [) = 0 then [) 0. Note that if the space is positive semi-denite
then if [) is a zero norm state and [) is any state we must have [) = 0, otherwise [)+[)
has negative norm for some . Hence the zero norm states form a subspace and furthermore it
is then easy to verify that

1
[
2
) =

1
[

2
) if [

1
) [
1
) , [

2
) [
2
) . (5.142)
85
So the identication of equivalent states is a well-dened operation which preserves scalar
products. This then gives a prescription for dening the space of physical states H
phys
= H
0
/
which satises the usual requirements of quantum mechanics. A further assumption is that the
zero norm states are BRS exact, i.e.
[) = 0 [) =

Q
BRS
[) for some [) . (5.143)
This ensures that H
phys
is formed by the cohomology classes of

Q
BRS
acting on the original
space H.
A further restriction is to consider just states with zero ghost number i.e.

Q
gh
[) = 0 , (5.144)
since physical states should satisfy this condition. Generally states with non zero ghost number
are BRS exact so they do not belong to H
phys
.
We now verify how this works for the subspace of single particle states. The eld b can
always be eliminated, so we do not have to consider a corresponding single particle state since
[b(k)) k

[A

(k)). The set of single particle states is spanned by

[A

(k)) , [c(k)) , [ c(k)) . (5.145)

The action of

Q
BRS
on these states is determined by the linear terms in the corresponding
action on the elds and must be of the form

Q
BRS
[A

(k)) = k

[c(k)) , (5.146a)

Q
BRS
[c(k)) = 0 , (5.146b)

Q
BRS
[ c(k)) = k

[A

(k)) , (5.146c)
for some and , using that

Q
BRS
increases the ghost charge by one, requiring Lorentz
invariance and furthermore assuming that

Q
BRS
maps single particle states to single particle
states. Clearly

Q
BRS
2
[ c(k)) = 0 k
2
= 0 . (5.147)
The action of the ghost number operator

Q
gh
, which counts the number of ghost elds, is
also given by

Q
gh
[c
a
(k)) = i[c
a
(k)) ,

Q
gh
[ c
a
(k)) = i[ c
a
(k)) , (5.148a)
c
a
(k)[

Q
gh
= ic
a
(k)[ , c
a
(k)[

Q
gh
= i c
a
(k)[ . (5.148b)
It is a interesting exercise to verify that

Q
gh
, although it has imaginary eigenvalues, is an
Hermitian operator. The ghost number may be dened as the integer eigenvalue of i

Q
gh
.
The scalar products which are determined by the propagators for the elds and must be
consistent with

Q
BRS
,

Q
gh
being hermitian. We assume a Lorentz invariant formalism as
appropriate for a linear covariant gauge, which requires that these single particle states have
the normalisation
A
a
(k

)[A
b
(k)) =

ab

k
,
k

k
:= (2)
d1
2k
0

d1
(

) , (5.149)
which yields negative norm for timelike components since
00
= 1, and, for the ghost elds,
c
a
(k

)[c
b
(k)) =
ab

k
. (5.150)
All other single particle scalar products are zero, in particular
c
a
(k

)[c
b
(k)) = 0 , (5.151)
so [c
a
(k)) is a zero norm state. By considering c

(k

)[

Q
BRS
[A

(k)) we must have =

.
To obtain a consistent physical theory it is necessary to reduce the theory from one dened
on space of states including those with negative norm H to a space of physical states which have
86
positive norm. H
phys
, the physical Hilbert
43
space, must contain only positive norm states
and it must also be invariant under time evolution. Since H
phys
has to be dened in terms of
H then this denition must not depend on how it is set up at any initial time.
The subspace of H
0
formed by single particle states that are annihilated by

Q
BRS
is then
determined by the basis

[A

(k)) , k = 0 , [c(k)) . (5.152)

With the equivalence
[) [) +

Q
BRS
[) , (5.153)
we must have

[A

(k)) (

+k

)[A

(k)) , [c) 0 . (5.154)

Hence H
phys
just consists of all states

[A

(k)), subject to
k = 0 ,

+k

. (5.155)
Note that in d-dimensions, both of these conditions remove one degree of freedom, so that
has just d 2 degrees of freedom. The physical states here just have zero ghost number.
Let us check the norm of these states:

(k)[A

(k

))

kk
, (5.156)
where
kk
is the standard Kronecker
44
delta, as far as the momentum is concerned. Note
that

= [
0
[
2
+[[
2
, (5.157)
and also
0 = k =
0
k
0
+

k =
0
[

k[ +

k , (5.158)
so that

= [[
2

1
[

k[
2
[

k[
2
0 . (5.159)
This is zero only for

k and then

, and for these states the equivalence implies

[A

(k)) 0 . (5.160)
5.8 Renormalisation of Gauge Theories
Gauge theories are renormalisable, with the degree of divergence given by
d = 4 E
A
E
gh
, (5.161)
where E
A
and E
gh
denotes the number of external gauge and ghost lines, respectively. Any
Feynman diagram with D 0 is expected to generate divergences that have to be cancelled.
The requirement of renormalisability is that

. . . A

(x) . . . c(y) . . . c(z) . . . b(w) . . .

_
, (5.162)
can be made nite, for arbitrary numbers of eld, by adding counterterms to the original
Lagrangian,
L L +L
c.t.
, (5.163)
in a fashion similar to previously. What counterterms are necessary is determined by D and
also by the form chosen for the gauge xing term. For the Lorentz invariant linear gauge
F(A) =

and using dimensional regularisation L

c.t.
contains all contributions formed
from local functions of the elds and their derivatives of dimension four, with the following
requirements:
43
Hilbert, David (1862-1943)
44
Kronecker, Leopold (1823-1891)
87
(i) Lorentz invariance,
(ii) BRS invariance,
(iii) conservation of ghost number.
Because of requirement (iii) each diagram has a c external line for every external c line so
that E
gh
must be event. The crucial cases for which D 0 are then for E
A
= 2, 3, 4, E
gh
= 0
(E
A
= 1 is absent because of Lorentz invariance), which require counterterms involving
A
2
, A
3
, A
4
and E
gh
= 2 and E
A
= 0, 1, for which the counterterms involve c c, c cA. For the
linear covariant gauge the cases E
gh
= E
A
= 2 and E
gh
= 4 do not require corresponding
counterterms as when there are ghost lines D is reduced by one and the associated Feynman
integrals do not diverge. This is because the Feynman rules for this simple gauge xing require
that for each c cA vertex the contribution contains just the momentum for the c line at this
vertex. For two such vertices which are part of a loop only one of the momentum factors is
involved in the loop integration rather than two which would be expected according to the naive
power counting rules so the degree of divergence for the loop integral is reduced by one.
It is possible to consider non-Lorentz invariant or non linear gauge xing conditions, but
in general this makes the renormalisation analysis more complicated. Requirement (i) ensures
that all counterterms are Lorentz scalars.
With d = 4 the starting theory is written as
L
q
=

1
g
2
_

1
4
F

(gA) F

(gA) s (gA, g c, gb)

_
, (5.164)
where we dene
F

(A) =

+ [A

, A

] , (A, c, b) = c

+
1
2
c b , (5.165)
and now, because of the rescaling of the elds by g,
sA

= D

(gA)c =

c +g[A

, c] , sc =
1
2
g[c, c] , s c = b , sb = 0 . (5.166)
The overall factor

, involving an arbitrary mass scale , has been introduced to ensure L

q
has
dimension d while A

, c, c have dimension 1, b dimension 2 while g is dimensionless. Furthermore

the elds have been rescaled by g to ensure that the quadratic terms are independent of g. As
has been discussed this is invariant under BRS transformations sL
q
= 0i.
The counterterms also have to be compatible with BRS symmetry. This is fundamental since
the whole denition of the physical theory in terms of states with positive norm depends on
the conserved BRS charge

Q
BRS
as a nilpotent operator. Dimensional regularisation preserves
BRS symmetry since it is valid for any d. To see the implications we dene
L
q,0
= L
q
+L
c.t.
, (5.167)
where we must require
sL
q,0
= 0 . (5.168)
However although we must have s
2
= 0 the action of s on the elds may be modied. L
q,0
contains various divergent constants, corresponding to poles in , which are traditionally labelled
Z. When Z = 1 in each case L
q,0
reduces to L
q
. In general for each Z there is an expansion
Z = 1 +

n=1
z
n
(g, )

n
, (5.169)
where the poles generate the necessary counterterms. The claim is that it is sucient in order
to get a nite result for Feynman integrals to any order to take
L
q,0
=

Z
g
4g
2
F

(gZ

A) F

(gZ

A) +Z

(gZ

A)c +b

+
1
2
b b
_
,
(5.170)
88
for suitable Z
g
, Z

, Z

. In this expression for L

q,0
we have assumed there are no divergences
associated with the b eld, it only contributes to one-particle reducible diagrams which do not
require separate counterterms to ensure niteness; hence there are no Z factors in the terms b
2
and bA. The action of s is modied to ensure sL
q,0
= 0 and also s
2
= 0,
sA

= D

(gZ

A)c , sc =
1
2
gZ

[c, c] , s c =
1
Z

b , sb = 0 . (5.171)
In consequence BRS symmetry remains valid for the nite theory obtained after regularisation
since
L
q,0
=

Z
g
4g
2
F

(gZ

A) F

(gZ

A) Z

s (A, c, b)
_
. (5.172)
In L
q,0
b can be eliminated by setting b =

/. It is important to note that there just

three Zs present in L
q,0
which can be used to cancel divergencies whereas there are ve dierent
cases for dierent numbers of external A and ghost lines for which there are divergent graphs.
Also the form of the counterterms which are contained in L
q,0
are not the most general Lorentz
invariant expressions either. Thus BRS invariance must constrain the dierent divergencies in
order for them to be cancelled by the allowed counterterms. Now let us illustrate, without
doing any calculations of Feynman integrals, how this works out at one loop. The results are
expressed in terms of the group theory constant C dened by
f
acd
f
bcd
= C
ab
, (5.173)
where for SU(N) C = N.
For E
A
= 2 and no external ghosts
L
q,0,AA
=
1
2
_
Z
g
Z

) +

/
_
. The one loop graphs are
, ,
,
, ,

A
A
A
A

We obtain the result

Z
g
Z

2
= 1 +
g
2
C
16
2
_
5
3
+
1
2
(1 )
_
2

, (5.174)
where C is some group constant, which is equal to N for SU(N).
For E
A
= 3 L
q,0,AAA
=

Z
g
Z

[A

, A

]. The one loop graphs are

89
These give
Z
g
Z

3
= 1 +
g
2
C
16
2
_
2
3
+
3
4
(1 )
_
2

. (5.175)
For E
A
= 4 L
q,0,AAAA
=
1
4
Z
g
Z

4
[A

, A

][A

, A

] and we have the one loop diagrams

These contribute
Z
g
Z

4
= 1 +
g
2
C
16
2
_

1
3
+ (1 )
_
2

. (5.176)
For two external ghost lines E
gh
= 2, E
A
= 0 L
q,0, cc
=

c
2
c. We get
Z

= 1 +
g
2
C
16
2
_
1
2
+
1
4
(1 )
_
2

. (5.177)
Finally for E
gh
= 2, E
A
= 1 L
q,0, ccA
=

[A

, c]. We get in this case

Z

= 1 +
g
2
C
16
2
_

1
2
+
1
2
(1 )
_
2

. (5.178)
These give ve independent results for three quantities, one can use this to check for con-
sistency and obtain the result
Z

= 1 +
g
2
C
16
2
_
1 +
1
4
(1 )
_
2

, (5.179)
and also
Z
g
= 1 +
g
2
C
16
2

11
3

2

. (5.180)
This is a very important result; Z
g
is independent of .
The Lagrangian L
q,0
including counterterms is the bare Lagrangian and can be written
in terms of bare elds and couplings in the form
L
q,0
=
1
4g
2
0
F

(g
0
A
0
) F

(g
0
A
0
) + c
0

(g
0
A
0
)c
0
+b
0

A
0
+
1
2

0
b
0
b
0
, (5.181)
where we dene
g
2
0
=
g
2
Z
g

,
0
= Z
g
Z
2

, (5.182)
90
and also for the bare elds
A
0
=

2
_
Z
g
Z

A, c
0
=

2
_
Z
g
Z

c , c
0
=

2
Z

_
Z
g
Z

c , b
0
=

2
1
_
Z
g
Z

b .
(5.183)
The denition of c
0
, c
0
is somewhat arbitrary since we may have c
0
c
0
while c
0
c
0
/.
With the above choice then the action of s in terms of the bare elds becomes
sc
0
=
1
2
g
0
[c
0
, c
0
] , sA
0
= D

(g
0
A
0
)c
0
, s c
0
= b
0
, sb
0
= 0 , (5.184)
which is the essentially the same as before, and
L
q,0
=
1
4g
2
0
F

(g
0
A
0
)F

(g
0
A
0
)s
0

0
(A
0
, c
0
, b
0
) ,
0
(A
0
, c
0
, b
0
) = c
0

A
0
+
1
2

0
c
0
b
0
.
(5.185)
From the one loop result for Z
g
1
g
2
0
=

_
1
g
2
+
C
16
2

11
3

2

_
. (5.186)
Now if you follow the procedures described earlier the -function is dened by

dg
d

g
0
=
1
2
g +(g) . (5.187)
Dierentiating g
0
2
with respect to then gives

_
1
g
2
+
C
16
2

11
3

2

_
=
_

1
2
g +(g)
_

g
1
g
2
, (5.188)
which implies that to this order
(g) =
11
3
g
3
C
16
2
< 0 . (5.189)
The minus sign indicates asymptotic freedom.
Now of course to calculate these things can be a mess, because the number of indices oating
around is quite large, one has to do a lot of contractions, and it is quite a non trivial exercise
unless one has previous experience. According to some historical recollections there were some
initial confusions as to signs by those who rst published a clear result for this calculation of
the -function, Gross and Wilczek and separately Politzer.
45
5.8.1 Calculation of (g) at One Loop
We now describe a more simplied calculation which leads to the same result for (g) at one
loop. The crucial step is to nd a way of calculation Z
g
directly without it being in combination
with other renormalisation constants which require a separate calculation.
To achieve this the quantum gauge eld /

is expanded about a xed classical background

A

, so that
/

= A

+ga

, (5.190)
where the functional integral is reduced to one over a

. In general to obtain a perturbation

expansion it is not necessary to expand around zero eld; in general relativity, for example, it
is usual to expand around the non zero metric for at space. Note that for the eld strength
T

= F

(/) = F

+g(D

) +g
2
[a

, a

] , (5.191)
45
Gross, David Jonathan (1941-), Wilczek, Frank Anthony (1951-), Politzer, Hugh David (1949-), Nobel Prizes
2004
91
for F

= F

(A) and we have dened

D

+ [A

, a

] , (5.192)
which is the covariant derivative for the background gauge eld A

.
The initial Lagrangian is
L =
1
4g
2
T

, (5.193)
which is invariant under gauge transformations on /

,
/

/
g

= g
1
/

g +g
1

g , T

T
g

= g
1
T

g . (5.194)
Because of the split into A

and a

this also implies invariance under

A

g
1
A

g +g
1

g , a

g
1
a

g , (5.195)
which is called a background gauge transformation. We claim that it is possible to maintain
background gauge invariance in the full quantum theory although it is still necessary to intro-
duce gauge xing term. In the functional integral we require
d[/] = d[a] , d[a] = d[g
1
ag] . (5.196)
The quantum action, eliminating b, is then given by
L
q
=
1
4g
2
T

1
2
F(a) F(a) + c F

c , (5.197)
where we have set = 1 for convenience. F(a) is the gauge xing condition and T

c =
D

c +g[a

, c].
That is so far standard, but the crucial idea is that under the background gauge transfor-
mation we may choose F(a) so that it is covariant, F(g
1
ag) = g
1
F(a)g, and so F(a) F(a)
is invariant. To achieve this we take
f(a) = D

. (5.198)
Then L
q
is invariant under background gauge transformations where the quantum eld a

transforms homogeneously. Nevertheless quantum gauge transformations which act only on the
dynamical eld a

, with A

xed,
a

a
g

= g
1
a

g +g
1

g +
1
g
_
g
1
A

g A

_
, (5.199)
and are a symmetry of the classical Lagrangian, are gauge xed are not a symmetry of L
q
. Note
that this expression for the gauge xing term reduces to the standard linear covariant choice
when A

= 0.
Dening S
q
[a, A, c, c] =
_
d
d
xL
q
then letting
Z[A] =
_
d[a]d[c]d[ c] e
iS
q
[a,A, c,c]
, (5.200)
we must have Z[A] = Z[A
g
]. We also dene
Z[A] = e
iW[A]
. (5.201)
Now in order to get a perturbative expansion, we expand S
q
in powers of a

; at lowest
order, at one loop, it is sucient to restrict the expansion to just the quadratic terms in a

, c, c.
Higher loop calculations involve the cubic and quartic terms as interactions. Hence we write
S
q
[a, c, c] =
1
g
2
S[A] +
1
g
_
d
d
x D

+S
a
+S
gh
+O(a
3
, cca) , (5.202)
92
where S[A] is just the basic action for the background eld, namely
S[A] =
1
4
_
d
d
x F

. (5.203)
For the quadratic terms it is easy to see what the ghost contribution is,
S
gh
=
_
d
d
x c D
2
c , (5.204)
where as before D

is the background covariant derivative. The result for S

a
which is quadratic
in a is more involved. Using the expansion of T

we get
S
a
=
1
2
_
d
d
x
_
D

(D

) [a

, a

] F

_
=
1
2
_
d
d
x a

_
D
2
a

+ [F

, a

] +D

_
, (5.205)
integrating by parts and using [a

, a

] F

= a

[F

, a

]. Noting also [D

, D

]a

= [F

, a

]
we get
S
a
=
1
2
_
d
d
x a

_
D
2
a

+ 2[F

, a

]
_
=:
1
2
_
d
d
x a

, (5.206)
where we dened the operator

= D
2

2F

a
T
a
, (5.207)
where T
a
are the generators of the Lie group in the adjoint representation.
Let us suppose that D

= 0, so in this case there is no linear term.

46
Then the one loop
approximation is obtained by
Z[A] = e
iW[A]
= e
i
g
2
S[A]
_
d[a]d[c]d[ c] e
i(S
a
+S
gh
)
, (5.209)
where S
a
, S
gh
are just the quadratic terms in the expansion given above. The Gaussian func-
tional integrals are readily evaluated in terms of the determinants of the relevant dierential
operators, giving
Z[A] = e
i
g
2
S[A]
det(D
2
)
(det )
1
2
, W[A] =
1
g
2
S[A] +
1
i
log det(D
2
)
1
2i
log det . (5.210)
It is important to recognise that under background gauge transformations D
2
g
1
D
2
g and
g
1
g so the determinants are gauge invariant. The determinants may be normalised so
that det(D
2
) = det = 1 when A

= 0 so that W[0] = 0.
The aim is now to calculate the divergent parts, or poles as = 4 d 0, of these two
determinants. The answers are given by
log det(D
2
)
1
3
2

C
16
2
iS[A] , log det =
20
3
2

C
16
2
iS[A] . (5.211)
The divergent terms can only involve S[A] since this is the unique gauge invariant quantity of
dimension four. The group constant C is here dened by
tr(T
a
T
b
) = C
ab
. (5.212)
46
Alternatively let
e
iW[A,J]
=
Z
d[a]d[c]d[ c] e
iS
q
[a,A, c,c]+
R
d
d
x J

(x)a

(x)
, (5.208)
and eliminate J by requiring W[A, J]/J(x) = 0.
93
Hence the divergent part will then be given by
W[A]
1
g
2
S[A]
11
3
2

C
16
2
S[A] . (5.213)
The pole can be cancelled by replacing in the rst term
1
g
2

Z
g
g
2
, (5.214)
and then letting
Z
g
= 1 +
11
3
2

Cg
2
16
2
. (5.215)
This is of course the same as previously, this gives directly the one loop result for (g) with the
correct sign for asymptotic freedom.
In order to calculate the divergent parts of the determinants of the dierential operators
depending on the background eld A

it is possible to use methods which maintain gauge

covariance throughout and are valid for any smooth background. This would require some
digression so instead we do it by brute force, expanding in the background eld A

to quadratic
order. This is sucient to determine the coecient of the poles which are proportional to
S[A] and reduces the calculation to that for one loop Feynman integrals.
First we consider det(D
2
). Since D

+A
a
T
a
we may expand
D
2
=
2
+

, A
a
T
a
+A
a
T
a
A

b
T
b
. (5.216)
In general for determinants of an operator X +Y we may expand in Y by using
log det
_
X +Y
_
= log det X + log det
_
1 +X
1
Y
_
= log det X + tr log
_
1 +X
1
Y
_
= log det X + tr
_
X
1
Y
_

1
2
tr
_
X
1
Y X
1
Y
_
+. . . . (5.217)
Applying this to obtain an expansion in A

we get
log det(D
2
) = log det(
2
) Tr
_
(
2
)
1
(

, A
a
T
a
+A
a
T
a
A

b
T
b
)
_

1
2
Tr
_
(
2
)
1

, A
a
T
a
(
2
)
1

, A
b
T
b
_
+O(A
3
) . (5.218)
The traces Tr are functional traces but they also include a conventional trace over group indices.
The functional trace can be discussed in various ways, one approach is to introduce bases
[x) and [k) such that
_
d
d
x[x)x[ =
_
d
d
k
(2)
d
[k)k[ = 1 , x[k) = e
ikx
. (5.219)
and

[k) = ik

[k) . (5.220)
We then have
x[(
2
)
1
[y) =
1
(2)
d
_
d
d
k
1
k
2
e
ik(xy)
. (5.221)
If X, Y are operators satisfying
X[x) = X(x)[x) , Y [x) = Y (x)[x) , (5.222)
then, since by denition Tr(O) =
_
d
d
xx[O[x),
Tr
_
(
2
)
1
X
_
=
_
d
d
x x[(
2
)
1
[x) X(x) =
1
(2)
d
_
d
d
k
1
k
2
_
d
d
x X(x) , (5.223)
94
and, introducing integrations over [x)x[ and also [y)y[,
Tr
_
(
2
)
1
X(
2
)
1
Y
_
=
_
d
d
x
_
d
d
y X(x)Y (y) x[(
2
)
1
[y)y[(
2
)
1
[x)
=
_
d
d
x
_
d
d
y X(x)Y (y)
1
(2)
2d
_
d
d
k d
d
k

e
i(kk

)(xy)
1
k
2
k
2
=
1
(2)
2d
_
d
d
p

X(p)

Y (p)
_
d
d
k
1
k
2
(k p)
2
, (5.224)
for

X(p) =
_
d
d
x e
ipx
X(x) ,

Y (p) =
_
d
d
x e
ipx
Y (x) . (5.225)
Applying these results to the calculation of log det(D
2
) then since tr(T
a
) = 0 there is no
linear term in A

. For the O(A

2
) term in the rst line this approach gives
Tr
_
(
2
)
1
A
a
T
a
A

b
T
b
_
=
_
d
d
x x[(
2
)
1
[x) tr
_
A
a
(x)T
a
A

b
(x)T
b
_
=
_
d
d
x tr
_
A
a
(x)T
a
A

b
(x)T
b
_
_
d
d
k
(2)
d
1
k
2
. (5.226)
This result may be represented by the Feynman graph, where A
a
and A
b
are attached to
the external lines,
,

@
@

@
@
k
A
a
A
b
For the remaining contribution we can calculate the trace in a similar fashion as above

1
2
Tr
_
(
2
)
1

, A
a
T
a
(
2
)
1

, A
b
T
b
_
=
1
2
_
d
d
x
_
d
d
y tr
_
A
a
(x)T
a
A
b
(y)T
b
_

_
y[

(
2
)
1

[x)x[(
2
)
1
[y) +y[

(
2
)
1
[x)x[

(
2
)
1
[y)
+n[(
2
)
1

[x)x[(
2
)
1

[y) +y[(
2
)
1
[x)x[

(
2
)
1

[y)
_
=
1
2
_
d
d
x
_
d
d
y tr
_
A
a
(x)T
a
A
b
(y)T
b
_
_
d
d
k
(2)
d
_
d
d
k

(2)
d
(k +k

(k +k

(k
2
i) (k
2
i)
e
i(kk

)(xy)
,
(5.227)
adopting now the usual i prescription in the denominators. The result is a Feynman integral
corresponding to the diagram with two external vector lines
A
a
, ,

A
A
A
A

-
p
-

k
k p
A
b
In terms of the Fourier transform

A
a
(p) =
_
d
d
x e
ipx
A
a
(x) we then have, letting
k

= k p,
log det
_
D
2
)/(
2
)
_
=
i
(2)
d
_
d
d
p I

(p) tr
_

A
a
(p)T
a

A
b
(p)T
b
_
+O(A
3
) , (5.228)
95
where
I

(p) =
1
2
1
(2)
d
i
_
d
d
k
(2k p)

(2k p)

(k
2
i)((k p)
2
i)

1
(2)
d
i
_
d
d
k
1
k
2
i
. (5.229)
This has the crucial property p

(p) = 0 which may be obtained from p(2kp) = k

2
(kp)
2
and using translation invariance of the integration. Furthermore using (2k p)
2
= 2(k p)
2
+
2k
2
p
2
we have

(p) = p
2
I(p) (d 2)
1
(2)
d
i
_
d
d
k
1
k
2
i
, (5.230)
where
I(p) =
1
2
1
(2)
d
i
_
d
d
k
1
(k
2
i)((k p)
2
i)
. (5.231)
Using dimensional regularisation the last term in I

(p) is zero
47
and hence
I

(p) =
1
d 1
_

p
2
p

_
I(p) . (5.233)
I(p) is a standard Feynman integral which has already been calculated. With m
2
= 0 this one
loop integral is given just in terms of Gamma functions
I(p) = (2
1
2
d)
(
1
2
d 1)
2
(d 2)
(p
2
)
1
2
d2
2(4)
1
2
d

1

1
16
2
. (5.234)
From the divergent part of I(p) we nd for I

(p)
48
I

(p)
1

1
3
1
16
2
(

p
2
p

) . (5.237)
With this result and
tr(T
a
T
b
) = C
ab
(5.238)
we get, to O(A
2
),
log det
_
D
2
/(
2
)
_

i
3
C
16
2
_
d
d
p
(2)
d
(

p
2
p

)

A
a
(p)

A
b
(p)
=
i
6
C
16
2
_
d
d
p
(2)
d
_
p

a
(p) p

b
(p)
__
p

A
a
(p) p

A
a
(p)
_
=
i
6
C
16
2
_
d
d
x
_

a
(x)

a
(x)
__

A
a
(x)

A
a
(x)
_
,
(5.239)
47
There are various ways to see that this must be the case; remember the integral
i
Z
d
d
k
(2)
d
1
k
2
+ m
2
=
1
(4)
d
2

1
d
2

(m
2
)
d
2
1
0 as m
2
0 , (5.232)
assuming d is analytically continued to d > 2.
48
A direct calculation of I

(p), using 1/(k

2
i) = i
R

0
de
i(k
2
i)
, is given by
I

(p) = i
2
Z

0
d
1
Z

0
d
2
1
2(2)
d
i
Z
d
d
k (2k p)

(2k p)

e
i
1
(k
2
i)i
2
((kp)
2
i)
= i
2
Z

0
d
1
Z

0
d
2
e
i

1

1
+
2
p
2

1
2(2)
d
i
Z
d
d
k

2k

1

2

1
+
2
p

2k

1

2

1
+
2
p

e
i(
1
+
2
)(k
2
i)
, (5.235)
with the usual trick of completing the square. Under integration k

k
2

/d while k

0. Then
carrying out the k

integration and letting

1
i
1
,
2
i
2
, assuming p
2
> 0, we have
I

(p) =
1
2(4)
1
2
d
Z

0
d
1
Z

0
d
2
e

1
+
2
p
2
1
(
1
+
2
)
1
2
d

1
+
2

1

2

1
+
2

2
p

=
1
(4)
1
2
d
(1
1
2
d)(
1
2
d)
2
(d)
(p
2
)
1
2
d2
`

p
2
p

. (5.236)
96
where the Fourier transform has been inverted back to position space. This result has a
unique gauge invariant completion giving
log det
_
D
2
/(
2
)
_

i
6
C
16
2
_
d
d
x F

a
(x) F
a
(x) =
2
3
C
16
2
iS[A] . (5.240)
The divergent part determinant of the operator acting on vector elds can be found in a
similar fashion. With the expansion

=
_

2
+

, A
a
T
a
+A
a
T
a
A

b
T
b
_

2F

a
T
a
, (5.241)
then
log det
_
/(
2
)1
_
= d Tr
_
(
2
)
1
A
a
T
a
A

b
T
b
_

1
2
d Tr((
2
)
1

, A
a
T
a
(
2
)
1

, A
b
T
b
_
2 Tr
_
(
2
)
1
F

a
T
a
(
2
)
1
F

b
T
b
_
+O(A
3
) , (5.242)
where we keep only the terms which are non zero to this order. The factors d arise from a trace
over Lorentz indices. For the nal term following the same methods as earlier

1
2
Tr
_
(
2
)
1
F

a
T
a
(
2
)
1
F

b
T
b
_
=
i
(2)
d
_
d
d
p I(p) tr
_

a
(p)T
a

F
b
(p)T
b
_
,
(5.243)
noting that F

a
F

b
= F

a
F
b
. The -pole in I(p) then gives

1
2
Tr
_
(
2
)
1
F

a
T
a
(
2
)
1
F

b
T
b
_

i

C
16
2
_
d
d
x F

a
(x) F
a
(x) . (5.244)
Combining this with the contribution from the rst two terms which is the same as that calcu-
lated earlier, apart from the additional factor of d, gives
log det
_
/(
2
)1
_

_
2
3
4
_
C
16
2
_
d
d
x F

a
(x) F
a
(x) =
40
3
C
16
2
iS[A] . (5.245)
97
6 Fermion Currents and Anomalies
The assumptions which lead to Noethers theorem and associated Ward-identities as de-
scribed in section 3.2 may be violated in quantum eld theories because the process of regular-
isation and renormlisation may violate some of the symmetries of the classical Lagrangian.
This is the case when the symmetries depend on a particular spacetime dimension so that they
are broken by dimensional regularisation. In this case there may be anomalies when classical
symmetries are not valid in the quantum eld theories. In particular such issues arise with
Dirac elds.
6.1 Example: Symmetries of the Dirac Lagrangian
Let us illustrate these issues by considering fermions; we have a Lagrangian
L =

( +m) , (6.1)
where the matrices satisfy

= 2

. (6.2)
The denition of

can be extended to any number of dimensions d. The free Lagrangian has

a U(1)
V
symmetry under the transformation
e
i
,



e
i
. (6.3)
The change in the action for an x dependent (x) is

S[,

] = i
_
d
d
x

(x)

(x)

(x) , (6.4)
which gives the conserved current
j

= i

. (6.5)
This is a conserved current under the Dirac equations
( +m) = 0 ,

+m
_
= 0 . (6.6)
We can also consider an axial U(1)
A
symmetry under a transformation
e
i
5
,



e
i
5
, (6.7)
where
2
5
= 1,

5
=
5
and

,
5
= 0 for all matrices. It is easy to see that in the massless
case m = 0,

L = 0 . (6.8)
If we follow the same procedure as before, we cannot extend
5
to d ,= 4 dimensions since it
involves the antisymmetric symbol

, the essential relation is the trace formula

49
tr
_

_
= 4i

. (6.11)
So now considering d = 4, the variation of the action is

S[,

] = i
_
d
4
x

(x)

5
, (6.12)
which gives the conserved axial current
j

5
= i

5
. (6.13)
49
For a demonstration of this note that
tr
`

= tr
`

, (6.9)
using cyclic symmetry of the trace and
5

5
. Using the standard properties of gamma matrices

= 2(

) +

. If

= d 1 the trace
formula gives
d

= 4

, (6.10)
using the antisymmetric properties of the -symbol.
98
6.2 Triangle Graphs
Now let us apply this to calculations and see how far we get in a case where there are potential
anomalies. Consider the non-trivial case for a correlation function for three currents
j

(x)j

(y)j

(z)) , (6.14)
and since the currents satisfy

= 0 we obtain from the Ward identity

(x)j

(y)j

(z)) = 0 . (6.15)
Taking the Fourier transform and factoring o a delta function:

(p, q, r)(2)
d

d
(p +q +r) =
_
d
d
xd
d
yd
d
z e
i(px+qy+rz)
j

(x)j

(y)j

(z)) . (6.16)
For free elds the Feynman diagrams are one loop,
-
p,

q,
?k +q
@I
@
k p
I
r,
+
-
p,

k p

q,
6
k +r
@R
@
k
I
r,
The propagators are given by

k
k
2
@I
@

Let us now assume m = 0 for that makes life slightly easier, then the Feynman rules will
give us

(p, q, r) = (i)
_
d
d
k
(2)
d
1
k
2
(k +q)
2
(k p)
2
tr
_

(k +q)

(k p)
_
+ (q, r, ) , (6.17)
where there is a sign associated with a Fermion loop. In principle we could evaluate this
integral, but that is not what we want. The integral has a degree of divergence D = d 3, so
that it would apparently not be convergent in four dimensions. However the leading term for
large k has the form
_
d
d
k
k k k
(k
2
)
3
= 0 . (6.18)
This reduces D by one.
The Ward identity asserts that
p

(p, q, r) = 0 , (6.19)
and there are corresponding identities involving q

and r

. To verify this identity we take the

contraction inside the integral and then use
k p (k p) = (k p) k
2
k (k p)
2
, (6.20)
99
which allows us to rewrite the integral as
p

(p, q, r) = (i)
_
d
d
k
(2)
d
1
(k +q)
2
tr
_

(k +q)

_
(k p)
(k p)
2

k
k
2
__
+ (q, r, ) . (6.21)
Now consider the shift in the integration so that k +q k, so that k p k +r, in the rst
term and this becomes
p

(p, q, r) = (i)
_
d
d
k
(2)
d
tr
_

k
k
2

(k +r)
(k +r)
2

(k +q)
(k +q)
2

k
k
2
_
+ (q, r, )
= 0 , (6.22)
using the cyclic property of the trace to show that the various terms cancel. Hence the Ward
identity is veried. Using dimensional regularisation all manipulations are justied but note
that in d = 4 dimensions, the change k+q k can generate surface terms. Crucially the U(1)
V
symmetry is valid for any d so the Ward identity is satised for the dimensionally regularised
theory. Requiring the Ward identity to be obeyed ensures a nite result when d = 4 as any
potential divergent terms fail to satisfy the identity.
Now consider a situation where there is one or three j
5
currents,
j

5
(x)j

(y)j

(z)) or j

5
(x)j

5
(y)j

5
(z)) . (6.23)
For the Fourier transform of j

5
(x)j

(y)j

(z)) in a similar fashion to the previous case, but

now setting d = 4 since
5
is only really dened in four dimensions,

(p, q, r) = (i)
_
d
4
k
(2)
4
1
k
2
(k +q)
2
(k p)
2
tr
_

(k +q)

5
(k p)
_
+ (q, r, ) . (6.24)
For j

5
j

5
j

5
) there are two additional
5
matrices in the trace but they cancel giving the same
result. With the same manipulations as in the previous example:
p

(p, q, r) = i
_
d
4
k
(2)
4
tr
_

k
k
2

(k +r)
(k +r)
2

5

(k +q)
(k +q)
2

k
k
2

5
_
+ (q, r, )
= 0 , (6.25)
where in the rst term we again shift the integration k + q q so that there is a similar
cancellation of terms as in the vector case. However the basic unregularised integral we are
dealing with is
_
d
4
k
k k
(k
2
)
2
, (6.26)
which is divergent for large k, so in the shift k k

= k + q can generate surface terms and

the above result can be modied.
For a well dened calculation, without ambiguity or the necessity of considering surface
terms under shifts of the integration momentum, it is necessary to regularise the integral.
Dimensional regularisation cannot be used so instead we change the propagator to

k
k
2
(k
2
) , (6.27)
where (k
2
) is a function satisfying (k
2
) = 1 for low k
2
and 0 for k
2
. This will
remove the large k divergence and the integral will have an additional factor in the integrand
(k
2
)((k +q)
2
) ((k p)
2
) . (6.28)
100
Hence p

(p, q, r) now becomes, after the integration shift k +q k, k p k +r in the

rst term,
p

(p, q, r) = i
_
d
4
k
(2)
4
tr
_

k
k
2

(k +r)
(k +r)
2

5
(k
2
)((k q)
2
)((k +r)
2
)

(k +q)
(k +q)
2

k
k
2

5
(k
2
)((k +q)
2
)((k p)
2
)
_
+ (q, r, ) . (6.29)
The whole expression can then be written in the form:
p

(p, q, r) = i
_
d
4
k
(2)
4
_
tr
_

k
k
2

(k +r)
(k +r)
2

5
_
(k
2
)((k +r)
2
)

_
((k q)
2
) ((k p)
2
)
_
tr
_

(k +q)
(k +q)
2

k
k
2

5
_
(k
2
)((k +q)
2
)

_
((k p)
2
) ((k r)
2
)
_
_
. (6.30)
For any nite k this is zero since then we can then take = 1. Let us on the other hand
consider what happens for k p, q, r, which gives the only possible non zero contributions to
the integral. For k large we may expand
((k q)
2
) ((k p)
2
)

(k
2
) 2k (q p), (6.31)
where

is the derivative of (k
2
) with respect to k
2
, of course. We can write by standard
properties of the Dirac matrices
tr
_

(k +r)
5
_
= 4i

. (6.32)
With these results the integral becomes
p

(p, q, r) = 4
_
d
4
k
(2)
4
1
(k
2
)
2
(k
2
)
2

(k
2
)

_
k

2k (qp)q

2k (pr)
_
, (6.33)
where it is possible to show that other contributions in the expansion of ((kq)
2
)((kp)
2
)
are unimportant. On integration k

1
4

k
2
. Also with Wick rotation d
4
k id
4
k and
then k
2
> 0,
p

(p, q, r) = 8i

_
d
4
k
(2)
4
1
k
2
(k
2
)
2

(k
2
) , (6.34)
using p = q r. Integrating over the angles, we can replace
d
4
k 2
2
k
3
dk =
2
d(k
2
)k
2
, (6.35)
so that, letting = k
2
, the essential integral we are dealing with becomes

_
0
d ()
2

() =
_
1
3
()
3
_

0
=
1
3
; (6.36)
which is independent of the detailed form of , save that (0) = 1 and that it vanishes at
innity. Finally we have obtained the non zero result
p

(p, q, r) =
1
6
2
i

, (6.37)
which constitutes an anomaly in that it disagrees with the naive expectation. There is a similar
result for the other identities, obtained under the simultaneous permutations ( , p
q r), since the correlation function is symmetric between all the external lines. Hence also
q

(p, q, r) =
1
6
2
i

, r

(p, q, r) =
1
6
2
i

. (6.38)
101
Other regularisation methods give the same answer, so long as the symmetry under interchange
of momenta and Lorentz indices for external lines is maintained, the method used above ensures
this since each Fermion propagator is regularised in the same way.
However in general there is a potential ambiguity in

, which results by letting

(p, q, r) +C i

(q r)

. (6.39)
where the last term will destroy the symmetry under permutations of the three external lines
although it is symmetric under q, r, . If
C =
1
6
2
, (6.40)
then including this extra piece we have
q

(p, q, r) = 0 , r

(p, q, r) = 0 , p

(p, q, r) =
1
2
2
i

. (6.41)
This is appropriate for j

5
(x)j

(y)j

(z)) since this need not be fully symmetric and the result
ensures there is no anomaly for the vector currents j

, j

.
Only the surface of anomalies have been touched here.
- END -
102